Diffusers Bot
commited on
Upload folder using huggingface_hub
Browse files- main/README.md +181 -205
- main/lpw_stable_diffusion_xl.py +1 -1
main/README.md
CHANGED
@@ -27,7 +27,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
|
|
27 |
| Multilingual Stable Diffusion | Stable Diffusion Pipeline that supports prompts in 50 different languages. | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline) | - | [Juan Carlos Piñeros](https://github.com/juancopi81) |
|
28 |
| GlueGen Stable Diffusion | Stable Diffusion Pipeline that supports prompts in different languages using GlueGen adapter. | [GlueGen Stable Diffusion](#gluegen-stable-diffusion-pipeline) | - | [Phạm Hồng Vinh](https://github.com/rootonchair) |
|
29 |
| Image to Image Inpainting Stable Diffusion | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | - | [Alex McKinney](https://github.com/vvvm23) |
|
30 |
-
| Text Based Inpainting Stable Diffusion | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting | [Text Based Inpainting Stable Diffusion](#
|
31 |
| Bit Diffusion | Diffusion on discrete data | [Bit Diffusion](#bit-diffusion) | - | [Stuti R.](https://github.com/kingstut) |
|
32 |
| K-Diffusion Stable Diffusion | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py) | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
|
33 |
| Checkpoint Merger Pipeline | Diffusion Pipeline that enables merging of saved model checkpoints | [Checkpoint Merger Pipeline](#checkpoint-merger-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
|
@@ -40,7 +40,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
|
|
40 |
| CLIP Guided Img2Img Stable Diffusion Pipeline | Doing CLIP guidance for image to image generation with Stable Diffusion | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion) | - | [Nipun Jindal](https://github.com/nipunjindal/) |
|
41 |
| TensorRT Stable Diffusion Text to Image Pipeline | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
|
42 |
| EDICT Image Editing Pipeline | Diffusion pipeline for text-guided image editing | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline) | - | [Joqsan Azocar](https://github.com/Joqsan) |
|
43 |
-
| Stable Diffusion RePaint | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.
|
44 |
| TensorRT Stable Diffusion Image to Image Pipeline | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
|
45 |
| Stable Diffusion IPEX Pipeline | Accelerate Stable Diffusion inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion on IPEX](#stable-diffusion-on-ipex) | - | [Yingjie Han](https://github.com/yingjie-han/) |
|
46 |
| CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |
|
@@ -192,10 +192,9 @@ prompt = "wooden boat"
|
|
192 |
init_image = load_image("https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/samples/images/2.jpg")
|
193 |
mask_image = load_image("https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/samples/masks/2.png")
|
194 |
|
195 |
-
image = pipe
|
196 |
|
197 |
make_image_grid([init_image, mask_image, image], rows=1, cols=3)
|
198 |
-
|
199 |
```
|
200 |
|
201 |
### Marigold Depth Estimation
|
@@ -223,7 +222,7 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
223 |
|
224 |
# (New) LCM version (faster speed)
|
225 |
pipe = DiffusionPipeline.from_pretrained(
|
226 |
-
"prs-eth/marigold-lcm-v1-0",
|
227 |
custom_pipeline="marigold_depth_estimation"
|
228 |
# torch_dtype=torch.float16, # (optional) Run with half-precision (16-bit float).
|
229 |
# variant="fp16", # (optional) Use with `torch_dtype=torch.float16`, to directly load fp16 checkpoint
|
@@ -366,7 +365,6 @@ guided_pipeline = DiffusionPipeline.from_pretrained(
|
|
366 |
custom_pipeline="clip_guided_stable_diffusion",
|
367 |
clip_model=clip_model,
|
368 |
feature_extractor=feature_extractor,
|
369 |
-
|
370 |
torch_dtype=torch.float16,
|
371 |
)
|
372 |
guided_pipeline.enable_attention_slicing()
|
@@ -394,7 +392,7 @@ for i, img in enumerate(images):
|
|
394 |
```
|
395 |
|
396 |
The `images` list contains a list of PIL images that can be saved locally or displayed directly in a google colab.
|
397 |
-
Generated images tend to be of higher
|
398 |
|
399 |
.
|
400 |
|
@@ -468,11 +466,9 @@ pipe.enable_attention_slicing()
|
|
468 |
|
469 |
|
470 |
### Text-to-Image
|
471 |
-
|
472 |
images = pipe.text2img("An astronaut riding a horse").images
|
473 |
|
474 |
### Image-to-Image
|
475 |
-
|
476 |
init_image = download_image("https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg")
|
477 |
|
478 |
prompt = "A fantasy landscape, trending on artstation"
|
@@ -480,7 +476,6 @@ prompt = "A fantasy landscape, trending on artstation"
|
|
480 |
images = pipe.img2img(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images
|
481 |
|
482 |
### Inpainting
|
483 |
-
|
484 |
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
|
485 |
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
|
486 |
init_image = download_image(img_url).resize((512, 512))
|
@@ -497,7 +492,7 @@ As shown above this one pipeline can run all both "text-to-image", "image-to-ima
|
|
497 |
Features of this custom pipeline:
|
498 |
|
499 |
- Input a prompt without the 77 token length limit.
|
500 |
-
- Includes tx2img, img2img
|
501 |
- Emphasize/weigh part of your prompt with parentheses as so: `a baby deer with (big eyes)`
|
502 |
- De-emphasize part of your prompt as so: `a [baby] deer with big eyes`
|
503 |
- Precisely weigh part of your prompt as so: `a baby deer with (big eyes:1.3)`
|
@@ -511,7 +506,7 @@ Prompt weighting equivalents:
|
|
511 |
|
512 |
You can run this custom pipeline as so:
|
513 |
|
514 |
-
####
|
515 |
|
516 |
```python
|
517 |
from diffusers import DiffusionPipeline
|
@@ -520,16 +515,14 @@ import torch
|
|
520 |
pipe = DiffusionPipeline.from_pretrained(
|
521 |
'hakurei/waifu-diffusion',
|
522 |
custom_pipeline="lpw_stable_diffusion",
|
523 |
-
|
524 |
torch_dtype=torch.float16
|
525 |
)
|
526 |
-
pipe=pipe.to("cuda")
|
527 |
|
528 |
prompt = "best_quality (1girl:1.3) bow bride brown_hair closed_mouth frilled_bow frilled_hair_tubes frills (full_body:1.3) fox_ear hair_bow hair_tubes happy hood japanese_clothes kimono long_sleeves red_bow smile solo tabi uchikake white_kimono wide_sleeves cherry_blossoms"
|
529 |
neg_prompt = "lowres, bad_anatomy, error_body, error_hair, error_arm, error_hands, bad_hands, error_fingers, bad_fingers, missing_fingers, error_legs, bad_legs, multiple_legs, missing_legs, error_lighting, error_shadow, error_reflection, text, error, extra_digit, fewer_digits, cropped, worst_quality, low_quality, normal_quality, jpeg_artifacts, signature, watermark, username, blurry"
|
530 |
|
531 |
-
pipe.text2img(prompt, negative_prompt=neg_prompt, width=512,height=512,max_embeddings_multiples=3).images[0]
|
532 |
-
|
533 |
```
|
534 |
|
535 |
#### onnxruntime
|
@@ -548,11 +541,10 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
548 |
prompt = "a photo of an astronaut riding a horse on mars, best quality"
|
549 |
neg_prompt = "lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
|
550 |
|
551 |
-
pipe.text2img(prompt,negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
|
552 |
-
|
553 |
```
|
554 |
|
555 |
-
|
556 |
|
557 |
### Speech to Image
|
558 |
|
@@ -587,7 +579,6 @@ diffuser_pipeline = DiffusionPipeline.from_pretrained(
|
|
587 |
custom_pipeline="speech_to_image_diffusion",
|
588 |
speech_model=model,
|
589 |
speech_processor=processor,
|
590 |
-
|
591 |
torch_dtype=torch.float16,
|
592 |
)
|
593 |
|
@@ -647,7 +638,6 @@ import torch
|
|
647 |
pipe = DiffusionPipeline.from_pretrained(
|
648 |
"CompVis/stable-diffusion-v1-4",
|
649 |
custom_pipeline="wildcard_stable_diffusion",
|
650 |
-
|
651 |
torch_dtype=torch.float16,
|
652 |
)
|
653 |
prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
|
@@ -707,7 +697,6 @@ for i in range(args.num_images):
|
|
707 |
images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
|
708 |
grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
|
709 |
tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
|
710 |
-
|
711 |
```
|
712 |
|
713 |
### Imagic Stable Diffusion
|
@@ -721,13 +710,14 @@ from io import BytesIO
|
|
721 |
import torch
|
722 |
import os
|
723 |
from diffusers import DiffusionPipeline, DDIMScheduler
|
|
|
724 |
has_cuda = torch.cuda.is_available()
|
725 |
device = torch.device('cpu' if not has_cuda else 'cuda')
|
726 |
pipe = DiffusionPipeline.from_pretrained(
|
727 |
"CompVis/stable-diffusion-v1-4",
|
728 |
-
|
729 |
custom_pipeline="imagic_stable_diffusion",
|
730 |
-
scheduler
|
731 |
).to(device)
|
732 |
generator = torch.Generator("cuda").manual_seed(0)
|
733 |
seed = 0
|
@@ -837,7 +827,7 @@ image.save('./seed_resize/seed_resize_{w}_{h}_image_compare.png'.format(w=width,
|
|
837 |
|
838 |
### Multilingual Stable Diffusion Pipeline
|
839 |
|
840 |
-
The following code can generate
|
841 |
|
842 |
```python
|
843 |
from PIL import Image
|
@@ -881,7 +871,6 @@ diffuser_pipeline = DiffusionPipeline.from_pretrained(
|
|
881 |
detection_pipeline=language_detection_pipeline,
|
882 |
translation_model=trans_model,
|
883 |
translation_tokenizer=trans_tokenizer,
|
884 |
-
|
885 |
torch_dtype=torch.float16,
|
886 |
)
|
887 |
|
@@ -905,9 +894,9 @@ This example produces the following images:
|
|
905 |
|
906 |
### GlueGen Stable Diffusion Pipeline
|
907 |
|
908 |
-
GlueGen is a minimal adapter that
|
909 |
|
910 |
-
Make sure you downloaded `gluenet_French_clip_overnorm_over3_noln.ckpt` for French (there are also pre-trained weights for Chinese, Italian, Japanese, Spanish or train your own) at [GlueGen's official repo](https://github.com/salesforce/GlueGen/tree/main)
|
911 |
|
912 |
```python
|
913 |
from PIL import Image
|
@@ -974,7 +963,6 @@ mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
|
|
974 |
pipe = DiffusionPipeline.from_pretrained(
|
975 |
"runwayml/stable-diffusion-inpainting",
|
976 |
custom_pipeline="img2img_inpainting",
|
977 |
-
|
978 |
torch_dtype=torch.float16
|
979 |
)
|
980 |
pipe = pipe.to("cuda")
|
@@ -1019,13 +1007,13 @@ image = pipe(image=image, text=text, prompt=prompt).images[0]
|
|
1019 |
|
1020 |
### Bit Diffusion
|
1021 |
|
1022 |
-
Based <https://arxiv.org/abs/2208.04202>, this is used for diffusion on discrete data - eg,
|
1023 |
|
1024 |
```python
|
1025 |
from diffusers import DiffusionPipeline
|
|
|
1026 |
pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="bit_diffusion")
|
1027 |
image = pipe().images[0]
|
1028 |
-
|
1029 |
```
|
1030 |
|
1031 |
### Stable Diffusion with K Diffusion
|
@@ -1091,37 +1079,36 @@ image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
|
|
1091 |
|
1092 |
### Checkpoint Merger Pipeline
|
1093 |
|
1094 |
-
Based on the AUTOMATIC1111/webui for checkpoint merging. This is a custom pipeline that merges
|
1095 |
|
1096 |
-
The checkpoint merging is currently memory intensive as it modifies the weights of a DiffusionPipeline object in place. Expect at least 13GB RAM
|
1097 |
-
on
|
1098 |
|
1099 |
Usage:-
|
1100 |
|
1101 |
```python
|
1102 |
from diffusers import DiffusionPipeline
|
1103 |
|
1104 |
-
#Return a CheckpointMergerPipeline class that allows you to merge checkpoints.
|
1105 |
-
#The checkpoint passed here is ignored. But still pass one of the checkpoints you plan to
|
1106 |
-
#merge for convenience
|
1107 |
pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="checkpoint_merger")
|
1108 |
|
1109 |
-
#There are multiple possible scenarios:
|
1110 |
-
#The pipeline with the merged checkpoints is returned in all the scenarios
|
1111 |
|
1112 |
-
#Compatible checkpoints a.k.a matched model_index.json files. Ignores the meta attributes in model_index.json during comparison.( attrs with _ as prefix )
|
1113 |
-
merged_pipe = pipe.merge(["CompVis/stable-diffusion-v1-4","CompVis/stable-diffusion-v1-2"], interp
|
1114 |
|
1115 |
-
#Incompatible checkpoints in model_index.json but merge might be possible. Use force
|
1116 |
-
merged_pipe_1 = pipe.merge(["CompVis/stable-diffusion-v1-4","hakurei/waifu-diffusion"], force
|
1117 |
|
1118 |
-
#Three checkpoint merging. Only "add_difference" method actually works on all three checkpoints. Using any other options will ignore the 3rd checkpoint.
|
1119 |
-
merged_pipe_2 = pipe.merge(["CompVis/stable-diffusion-v1-4","hakurei/waifu-diffusion","prompthero/openjourney"], force
|
1120 |
|
1121 |
prompt = "An astronaut riding a horse on Mars"
|
1122 |
|
1123 |
image = merged_pipe(prompt).images[0]
|
1124 |
-
|
1125 |
```
|
1126 |
|
1127 |
Some examples along with the merge details:
|
@@ -1132,7 +1119,7 @@ Some examples along with the merge details:
|
|
1132 |
|
1133 |
2. "hakurei/waifu-diffusion" + "prompthero/openjourney" ; Inverse Sigmoid interpolation; alpha = 0.8
|
1134 |
|
1135 |
-

|
1231 |
-
StableDiffusionImageVariationPipeline("lambdalabs/sd-image-variations-diffusers")
|
1232 |
|
1233 |
```python
|
1234 |
import torch
|
@@ -1269,7 +1256,7 @@ image.save("./shiba-inu.jpg")
|
|
1269 |
print(pipeline.decoder_pipe.__class__)
|
1270 |
# <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline'>
|
1271 |
|
1272 |
-
# this pipeline only
|
1273 |
# It is used to convert clip text embedding to clip image embedding.
|
1274 |
print(pipeline)
|
1275 |
# StableUnCLIPPipeline {
|
@@ -1329,10 +1316,10 @@ pipe.to(device)
|
|
1329 |
|
1330 |
start_prompt = "A photograph of an adult lion"
|
1331 |
end_prompt = "A photograph of a lion cub"
|
1332 |
-
#For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
|
1333 |
generator = torch.Generator(device=device).manual_seed(42)
|
1334 |
|
1335 |
-
output = pipe(start_prompt, end_prompt, steps
|
1336 |
|
1337 |
for i,image in enumerate(output.images):
|
1338 |
img.save('result%s.jpg' % i)
|
@@ -1367,10 +1354,10 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
1367 |
pipe.to(device)
|
1368 |
|
1369 |
images = [Image.open('./starry_night.jpg'), Image.open('./flowers.jpg')]
|
1370 |
-
#For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
|
1371 |
generator = torch.Generator(device=device).manual_seed(42)
|
1372 |
|
1373 |
-
output = pipe(image
|
1374 |
|
1375 |
for i,image in enumerate(output.images):
|
1376 |
image.save('starry_to_flowers_%s.jpg' % i)
|
@@ -1392,7 +1379,7 @@ The resulting images in order:-
|
|
1392 |
|
1393 |
### DDIM Noise Comparative Analysis Pipeline
|
1394 |
|
1395 |
-
#### **Research
|
1396 |
|
1397 |
The [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227) paper proposed an approach to answer the above question, which is their second contribution.
|
1398 |
The approach consists of the following steps:
|
@@ -1409,7 +1396,7 @@ import torch
|
|
1409 |
from PIL import Image
|
1410 |
import numpy as np
|
1411 |
|
1412 |
-
image_path = "path/to/your/image"
|
1413 |
image_pil = Image.open(image_path)
|
1414 |
image_name = image_path.split("/")[-1].split(".")[0]
|
1415 |
|
@@ -1448,6 +1435,7 @@ import torch
|
|
1448 |
from diffusers import DiffusionPipeline
|
1449 |
from PIL import Image
|
1450 |
from transformers import CLIPFeatureExtractor, CLIPModel
|
|
|
1451 |
feature_extractor = CLIPFeatureExtractor.from_pretrained(
|
1452 |
"laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
|
1453 |
)
|
@@ -1622,6 +1610,7 @@ import requests
|
|
1622 |
import torch
|
1623 |
from io import BytesIO
|
1624 |
from diffusers import StableDiffusionPipeline, RePaintScheduler
|
|
|
1625 |
def download_image(url):
|
1626 |
response = requests.get(url)
|
1627 |
return PIL.Image.open(BytesIO(response.content)).convert("RGB")
|
@@ -1679,7 +1668,7 @@ image.save('tensorrt_img2img_new_zealand_hills.png')
|
|
1679 |
```
|
1680 |
|
1681 |
### Stable Diffusion BoxDiff
|
1682 |
-
BoxDiff is a training-free method for controlled generation with bounding box coordinates. It
|
1683 |
```py
|
1684 |
import torch
|
1685 |
from PIL import Image, ImageDraw
|
@@ -1839,13 +1828,13 @@ Output Image
|
|
1839 |
|
1840 |
### Stable Diffusion on IPEX
|
1841 |
|
1842 |
-
This diffusion pipeline aims to
|
1843 |
|
1844 |
To use this pipeline, you need to:
|
1845 |
|
1846 |
1. Install [IPEX](https://github.com/intel/intel-extension-for-pytorch)
|
1847 |
|
1848 |
-
**Note:** For each PyTorch release, there is a corresponding release of the IPEX. Here is the mapping relationship. It is recommended to install
|
1849 |
|
1850 |
|PyTorch Version|IPEX Version|
|
1851 |
|--|--|
|
@@ -1864,26 +1853,26 @@ python -m pip install intel_extension_for_pytorch
|
|
1864 |
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
|
1865 |
```
|
1866 |
|
1867 |
-
2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX
|
1868 |
|
1869 |
**Note:** The setting of generated image height/width for `prepare_for_ipex()` should be same as the setting of pipeline inference.
|
1870 |
|
1871 |
```python
|
1872 |
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_ipex")
|
1873 |
# For Float32
|
1874 |
-
pipe.prepare_for_ipex(prompt, dtype=torch.float32, height=512, width=512) #value of image height/width should be consistent with the pipeline inference
|
1875 |
# For BFloat16
|
1876 |
-
pipe.prepare_for_ipex(prompt, dtype=torch.bfloat16, height=512, width=512) #value of image height/width should be consistent with the pipeline inference
|
1877 |
```
|
1878 |
|
1879 |
Then you can use the ipex pipeline in a similar way to the default stable diffusion pipeline.
|
1880 |
|
1881 |
```python
|
1882 |
# For Float32
|
1883 |
-
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
|
1884 |
# For BFloat16
|
1885 |
with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
1886 |
-
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
|
1887 |
```
|
1888 |
|
1889 |
The following code compares the performance of the original stable diffusion pipeline with the ipex-optimized pipeline.
|
@@ -1901,7 +1890,7 @@ def elapsed_time(pipeline, nb_pass=3, num_inference_steps=20):
|
|
1901 |
# warmup
|
1902 |
for _ in range(2):
|
1903 |
images = pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512).images
|
1904 |
-
#time evaluation
|
1905 |
start = time.time()
|
1906 |
for _ in range(nb_pass):
|
1907 |
pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512)
|
@@ -1922,7 +1911,7 @@ with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
|
1922 |
latency = elapsed_time(pipe)
|
1923 |
print("Latency of StableDiffusionIPEXPipeline--bf16", latency)
|
1924 |
latency = elapsed_time(pipe2)
|
1925 |
-
print("Latency of StableDiffusionPipeline--bf16",latency)
|
1926 |
|
1927 |
############## fp32 inference performance ###############
|
1928 |
|
@@ -1937,13 +1926,12 @@ pipe4 = StableDiffusionPipeline.from_pretrained(model_id)
|
|
1937 |
latency = elapsed_time(pipe3)
|
1938 |
print("Latency of StableDiffusionIPEXPipeline--fp32", latency)
|
1939 |
latency = elapsed_time(pipe4)
|
1940 |
-
print("Latency of StableDiffusionPipeline--fp32",latency)
|
1941 |
-
|
1942 |
```
|
1943 |
|
1944 |
### Stable Diffusion XL on IPEX
|
1945 |
|
1946 |
-
This diffusion pipeline aims to
|
1947 |
|
1948 |
To use this pipeline, you need to:
|
1949 |
|
@@ -1968,7 +1956,7 @@ python -m pip install intel_extension_for_pytorch
|
|
1968 |
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
|
1969 |
```
|
1970 |
|
1971 |
-
2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX
|
1972 |
|
1973 |
**Note:** The values of `height` and `width` used during preparation with `prepare_for_ipex()` should be the same when running inference with the prepared pipeline.
|
1974 |
|
@@ -2011,7 +1999,7 @@ def elapsed_time(pipeline, nb_pass=3, num_inference_steps=1):
|
|
2011 |
# warmup
|
2012 |
for _ in range(2):
|
2013 |
images = pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512, guidance_scale=0.0).images
|
2014 |
-
#time evaluation
|
2015 |
start = time.time()
|
2016 |
for _ in range(nb_pass):
|
2017 |
pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512, guidance_scale=0.0)
|
@@ -2047,8 +2035,7 @@ pipe4 = StableDiffusionXLPipeline.from_pretrained(model_id, low_cpu_mem_usage=Tr
|
|
2047 |
latency = elapsed_time(pipe3, num_inference_steps=steps)
|
2048 |
print("Latency of StableDiffusionXLPipelineIpex--fp32", latency, "s for total", steps, "steps")
|
2049 |
latency = elapsed_time(pipe4, num_inference_steps=steps)
|
2050 |
-
print("Latency of StableDiffusionXLPipeline--fp32",latency, "s for total", steps, "steps")
|
2051 |
-
|
2052 |
```
|
2053 |
|
2054 |
### CLIP Guided Images Mixing With Stable Diffusion
|
@@ -2061,7 +2048,7 @@ This approach is using (optional) CoCa model to avoid writing image description.
|
|
2061 |
|
2062 |
### Stable Diffusion XL Long Weighted Prompt Pipeline
|
2063 |
|
2064 |
-
This SDXL pipeline
|
2065 |
|
2066 |
You can provide both `prompt` and `prompt_2`. If only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline.
|
2067 |
|
@@ -2089,31 +2076,31 @@ pipe.to("cuda")
|
|
2089 |
t2i_images = pipe(
|
2090 |
prompt=prompt,
|
2091 |
negative_prompt=neg_prompt,
|
2092 |
-
).images
|
2093 |
|
2094 |
# img2img
|
2095 |
-
input_image = load_image("/path/to/local/image.png")
|
2096 |
i2i_images = pipe.img2img(
|
2097 |
prompt=prompt,
|
2098 |
negative_prompt=neg_prompt,
|
2099 |
image=input_image,
|
2100 |
-
strength=0.8,
|
2101 |
).images
|
2102 |
|
2103 |
# inpaint
|
2104 |
-
input_mask = load_image("/path/to/local/mask.png")
|
2105 |
inpaint_images = pipe.inpaint(
|
2106 |
prompt="photo of a cute (black) cat running on the grass" * 20,
|
2107 |
negative_prompt=neg_prompt,
|
2108 |
image=input_image,
|
2109 |
mask=input_mask,
|
2110 |
-
strength=0.6,
|
2111 |
).images
|
2112 |
|
2113 |
pipe.to("cpu")
|
2114 |
torch.cuda.empty_cache()
|
2115 |
|
2116 |
-
from IPython.display import display
|
2117 |
display(t2i_images[0])
|
2118 |
display(i2i_images[0])
|
2119 |
display(inpaint_images[0])
|
@@ -2153,9 +2140,9 @@ coca_model = open_clip.create_model('coca_ViT-L-14', pretrained='laion2B-s13B-b9
|
|
2153 |
coca_model.dtype = torch.float16
|
2154 |
coca_transform = open_clip.image_transform(
|
2155 |
coca_model.visual.image_size,
|
2156 |
-
is_train
|
2157 |
-
mean
|
2158 |
-
std
|
2159 |
)
|
2160 |
coca_tokenizer = SimpleTokenizer()
|
2161 |
|
@@ -2207,7 +2194,7 @@ This pipeline uses the Mixture. Refer to the [Mixture](https://arxiv.org/abs/230
|
|
2207 |
```python
|
2208 |
from diffusers import LMSDiscreteScheduler, DiffusionPipeline
|
2209 |
|
2210 |
-
#
|
2211 |
scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
|
2212 |
pipeline = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler, custom_pipeline="mixture_tiling")
|
2213 |
pipeline.to("cuda")
|
@@ -2248,7 +2235,6 @@ from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
|
|
2248 |
# Use the PNDMScheduler scheduler here instead
|
2249 |
scheduler = PNDMScheduler.from_pretrained("stabilityai/stable-diffusion-2-inpainting", subfolder="scheduler")
|
2250 |
|
2251 |
-
|
2252 |
pipe = StableDiffusionInpaintPipeline.from_pretrained("stabilityai/stable-diffusion-2-inpainting",
|
2253 |
custom_pipeline="stable_diffusion_tensorrt_inpaint",
|
2254 |
variant='fp16',
|
@@ -2287,7 +2273,7 @@ from diffusers.pipelines.pipeline_utils import Image2ImageRegion, Text2ImageRegi
|
|
2287 |
# Load and preprocess guide image
|
2288 |
iic_image = preprocess_image(Image.open("input_image.png").convert("RGB"))
|
2289 |
|
2290 |
-
#
|
2291 |
scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
|
2292 |
pipeline = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler).to("cuda:0", custom_pipeline="mixture_canvas")
|
2293 |
pipeline.to("cuda")
|
@@ -2298,7 +2284,7 @@ output = pipeline(
|
|
2298 |
canvas_width=352,
|
2299 |
regions=[
|
2300 |
Text2ImageRegion(0, 800, 0, 352, guidance_scale=8,
|
2301 |
-
prompt=f"best quality, masterpiece, WLOP, sakimichan, art contest winner on pixiv, 8K, intricate details, wet effects, rain drops, ethereal, mysterious, futuristic, UHD, HDR, cinematic lighting, in a beautiful forest, rainy day, award winning, trending on artstation, beautiful confident cheerful young woman, wearing a futuristic sleeveless dress, ultra beautiful detailed eyes, hyper-detailed face, complex, perfect, model,
|
2302 |
Image2ImageRegion(352-800, 352, 0, 352, reference_image=iic_image, strength=1.0),
|
2303 |
],
|
2304 |
num_inference_steps=100,
|
@@ -2317,22 +2303,19 @@ It is a simple and minimalist diffusion model.
|
|
2317 |
The following code shows how to use the IADB pipeline to generate images using a pretrained celebahq-256 model.
|
2318 |
|
2319 |
```python
|
2320 |
-
|
2321 |
pipeline_iadb = DiffusionPipeline.from_pretrained("thomasc4/iadb-celebahq-256", custom_pipeline='iadb')
|
2322 |
|
2323 |
pipeline_iadb = pipeline_iadb.to('cuda')
|
2324 |
|
2325 |
-
output = pipeline_iadb(batch_size=4,num_inference_steps=128)
|
2326 |
for i in range(len(output[0])):
|
2327 |
plt.imshow(output[0][i])
|
2328 |
plt.show()
|
2329 |
-
|
2330 |
```
|
2331 |
|
2332 |
Sampling with the IADB formulation is easy, and can be done in a few lines (the pipeline already implements it):
|
2333 |
|
2334 |
```python
|
2335 |
-
|
2336 |
def sample_iadb(model, x0, nb_step):
|
2337 |
x_alpha = x0
|
2338 |
for t in range(nb_step):
|
@@ -2343,13 +2326,11 @@ def sample_iadb(model, x0, nb_step):
|
|
2343 |
x_alpha = x_alpha + (alpha_next-alpha)*d
|
2344 |
|
2345 |
return x_alpha
|
2346 |
-
|
2347 |
```
|
2348 |
|
2349 |
The training loop is also straightforward:
|
2350 |
|
2351 |
```python
|
2352 |
-
|
2353 |
# Training loop
|
2354 |
while True:
|
2355 |
x0 = sample_noise()
|
@@ -2380,7 +2361,7 @@ import torch
|
|
2380 |
from pipeline_zero1to3 import Zero1to3StableDiffusionPipeline
|
2381 |
from diffusers.utils import load_image
|
2382 |
|
2383 |
-
model_id = "kxic/zero123-165000"
|
2384 |
|
2385 |
pipe = Zero1to3StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
|
2386 |
|
@@ -2401,9 +2382,9 @@ query_pose3 = [-55.0, 90.0, 0.0]
|
|
2401 |
# H, W = (256, 256) # H, W = (512, 512) # zero123 training is 256,256
|
2402 |
|
2403 |
# for batch input
|
2404 |
-
input_image1 = load_image("./demo/4_blackarm.png")
|
2405 |
-
input_image2 = load_image("./demo/8_motor.png")
|
2406 |
-
input_image3 = load_image("./demo/7_london.png")
|
2407 |
input_images = [input_image1, input_image2, input_image3]
|
2408 |
query_poses = [query_pose1, query_pose2, query_pose3]
|
2409 |
|
@@ -2434,7 +2415,6 @@ input_images = pre_images
|
|
2434 |
images = pipe(input_imgs=input_images, prompt_imgs=input_images, poses=query_poses, height=H, width=W,
|
2435 |
guidance_scale=3.0, num_images_per_prompt=num_images_per_prompt, num_inference_steps=50).images
|
2436 |
|
2437 |
-
|
2438 |
# save imgs
|
2439 |
log_dir = "logs"
|
2440 |
os.makedirs(log_dir, exist_ok=True)
|
@@ -2444,12 +2424,11 @@ for obj in range(bs):
|
|
2444 |
for idx in range(num_images_per_prompt):
|
2445 |
images[i].save(os.path.join(log_dir,f"obj{obj}_{idx}.jpg"))
|
2446 |
i += 1
|
2447 |
-
|
2448 |
```
|
2449 |
|
2450 |
### Stable Diffusion XL Reference
|
2451 |
|
2452 |
-
This pipeline uses the Reference
|
2453 |
|
2454 |
```py
|
2455 |
import torch
|
@@ -2457,6 +2436,7 @@ from PIL import Image
|
|
2457 |
from diffusers.utils import load_image
|
2458 |
from diffusers import DiffusionPipeline
|
2459 |
from diffusers.schedulers import UniPCMultistepScheduler
|
|
|
2460 |
input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
|
2461 |
|
2462 |
# pipe = DiffusionPipeline.from_pretrained(
|
@@ -2529,7 +2509,7 @@ from diffusers import DiffusionPipeline
|
|
2529 |
# load the pipeline
|
2530 |
# make sure you're logged in with `huggingface-cli login`
|
2531 |
model_id_or_path = "runwayml/stable-diffusion-v1-5"
|
2532 |
-
#can also be used with dreamlike-art/dreamlike-photoreal-2.0
|
2533 |
pipe = DiffusionPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16, custom_pipeline="pipeline_fabric").to("cuda")
|
2534 |
|
2535 |
# let's specify a prompt
|
@@ -2560,7 +2540,7 @@ torch.manual_seed(0)
|
|
2560 |
image = pipe(
|
2561 |
prompt=prompt,
|
2562 |
negative_prompt=negative_prompt,
|
2563 |
-
liked
|
2564 |
num_inference_steps=20,
|
2565 |
).images[0]
|
2566 |
|
@@ -2730,7 +2710,7 @@ pipe.to(torch_device="cuda", torch_dtype=torch.float32)
|
|
2730 |
```py
|
2731 |
prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
|
2732 |
|
2733 |
-
# Can be set to 1~50 steps. LCM
|
2734 |
num_inference_steps = 4
|
2735 |
|
2736 |
images = pipe(prompt=prompt, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
|
@@ -2762,9 +2742,9 @@ prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
|
|
2762 |
|
2763 |
input_image=Image.open("myimg.png")
|
2764 |
|
2765 |
-
strength = 0.5
|
2766 |
|
2767 |
-
# Can be set to 1~50 steps. LCM
|
2768 |
num_inference_steps = 4
|
2769 |
|
2770 |
images = pipe(prompt=prompt, image=input_image, strength=strength, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
|
@@ -2808,7 +2788,7 @@ images = pipe(
|
|
2808 |
guidance_scale=8.0,
|
2809 |
embedding_interpolation_type="lerp",
|
2810 |
latent_interpolation_type="slerp",
|
2811 |
-
process_batch_size=4,
|
2812 |
generator=torch.Generator(seed),
|
2813 |
)
|
2814 |
|
@@ -2827,7 +2807,7 @@ Two checkpoints are available for use:
|
|
2827 |
- [ldm3d-pano](https://huggingface.co/Intel/ldm3d-pano). This checkpoint enables the generation of panoramic images and requires the StableDiffusionLDM3DPipeline pipeline to be used.
|
2828 |
- [ldm3d-sr](https://huggingface.co/Intel/ldm3d-sr). This checkpoint enables the upscaling of RGB and depth images. Can be used in cascade after the original LDM3D pipeline using the StableDiffusionUpscaleLDM3DPipeline pipeline.
|
2829 |
|
2830 |
-
|
2831 |
from PIL import Image
|
2832 |
import os
|
2833 |
import torch
|
@@ -2838,11 +2818,11 @@ from diffusers import StableDiffusionLDM3DPipeline, DiffusionPipeline
|
|
2838 |
pipe_ldm3d = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
|
2839 |
pipe_ldm3d.to("cuda")
|
2840 |
|
2841 |
-
prompt =
|
2842 |
output = pipe_ldm3d(prompt)
|
2843 |
rgb_image, depth_image = output.rgb, output.depth
|
2844 |
-
rgb_image[0].save(
|
2845 |
-
depth_image[0].save(
|
2846 |
|
2847 |
# Upscale the previous output to a resolution of (1024, 1024)
|
2848 |
|
@@ -2850,19 +2830,19 @@ pipe_ldm3d_upscale = DiffusionPipeline.from_pretrained("Intel/ldm3d-sr", custom_
|
|
2850 |
|
2851 |
pipe_ldm3d_upscale.to("cuda")
|
2852 |
|
2853 |
-
low_res_img = Image.open(
|
2854 |
-
low_res_depth = Image.open(
|
2855 |
outputs = pipe_ldm3d_upscale(prompt="high quality high resolution uhd 4k image", rgb=low_res_img, depth=low_res_depth, num_inference_steps=50, target_res=[1024, 1024])
|
2856 |
|
2857 |
-
upscaled_rgb, upscaled_depth =outputs.rgb[0], outputs.depth[0]
|
2858 |
-
upscaled_rgb.save(
|
2859 |
-
upscaled_depth.save(
|
2860 |
-
|
2861 |
|
2862 |
### ControlNet + T2I Adapter Pipeline
|
2863 |
|
2864 |
-
This
|
2865 |
-
It receives `control_image` and `adapter_image`, as well as `controlnet_conditioning_scale` and `adapter_conditioning_scale`, for the ControlNet and Adapter modules, respectively. Whenever `adapter_conditioning_scale
|
2866 |
|
2867 |
```py
|
2868 |
import cv2
|
@@ -2925,7 +2905,6 @@ images = pipe(
|
|
2925 |
adapter_conditioning_scale=strength,
|
2926 |
).images
|
2927 |
images[0].save("controlnet_and_adapter.png")
|
2928 |
-
|
2929 |
```
|
2930 |
|
2931 |
### ControlNet + T2I Adapter + Inpainting Pipeline
|
@@ -2996,12 +2975,11 @@ images = pipe(
|
|
2996 |
strength=0.7,
|
2997 |
).images
|
2998 |
images[0].save("controlnet_and_adapter_inpaint.png")
|
2999 |
-
|
3000 |
```
|
3001 |
|
3002 |
### Regional Prompting Pipeline
|
3003 |
|
3004 |
-
This pipeline is a port of the [Regional Prompter extension](https://github.com/hako-mikan/sd-webui-regional-prompter) for [Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui) to diffusers
|
3005 |
This code implements a pipeline for the Stable Diffusion model, enabling the division of the canvas into multiple regions, with different prompts applicable to each region. Users can specify regions in two ways: using `Cols` and `Rows` modes for grid-like divisions, or the `Prompt` mode for regions calculated based on prompts.
|
3006 |
|
3007 |

|
@@ -3012,6 +2990,7 @@ This code implements a pipeline for the Stable Diffusion model, enabling the div
|
|
3012 |
|
3013 |
```py
|
3014 |
from examples.community.regional_prompting_stable_diffusion import RegionalPromptingStableDiffusionPipeline
|
|
|
3015 |
pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae=vae)
|
3016 |
|
3017 |
rp_args = {
|
@@ -3019,7 +2998,7 @@ rp_args = {
|
|
3019 |
"div": "1;1;1"
|
3020 |
}
|
3021 |
|
3022 |
-
prompt ="""
|
3023 |
green hair twintail BREAK
|
3024 |
red blouse BREAK
|
3025 |
blue skirt
|
@@ -3029,12 +3008,12 @@ images = pipe(
|
|
3029 |
prompt=prompt,
|
3030 |
negative_prompt=negative_prompt,
|
3031 |
guidance_scale=7.5,
|
3032 |
-
height
|
3033 |
-
width
|
3034 |
-
num_inference_steps
|
3035 |
-
num_images_per_prompt
|
3036 |
-
rp_args
|
3037 |
-
|
3038 |
|
3039 |
time = time.strftime(r"%Y%m%d%H%M%S")
|
3040 |
i = 1
|
@@ -3059,19 +3038,19 @@ blue skirt
|
|
3059 |
|
3060 |
### 2-Dimentional division
|
3061 |
|
3062 |
-
The prompt consists of instructions separated by the term `BREAK` and is assigned to different regions of a two-dimensional space. The image is initially split in the main splitting direction, which in this case is rows, due to the presence of a single semicolon`;`, dividing the space into an upper and a lower section. Additional sub-splitting is then applied, indicated by commas. The upper row is split into ratios of `2:1:1`, while the lower row is split into a ratio of `4:6`. Rows themselves are split in a `1:2` ratio. According to the reference image, the blue sky is designated as the first region, green hair as the second, the bookshelf as the third, and so on, in a sequence based on their position from the top left. The terrarium is placed on the desk in the fourth region, and the orange dress and sofa are in the fifth region, conforming to their respective splits.
|
3063 |
|
3064 |
-
```
|
3065 |
rp_args = {
|
3066 |
"mode":"rows",
|
3067 |
"div": "1,2,1,1;2,4,6"
|
3068 |
}
|
3069 |
|
3070 |
-
prompt ="""
|
3071 |
blue sky BREAK
|
3072 |
green hair BREAK
|
3073 |
book shelf BREAK
|
3074 |
-
terrarium on desk BREAK
|
3075 |
orange dress and sofa
|
3076 |
"""
|
3077 |
```
|
@@ -3080,10 +3059,10 @@ orange dress and sofa
|
|
3080 |
|
3081 |
### Prompt Mode
|
3082 |
|
3083 |
-
There are limitations to methods of specifying regions in advance. This is because specifying regions can be a hindrance when designating complex shapes or dynamic compositions. In the region specified by the prompt, the
|
3084 |
For further infomagen, see [here](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/main/prompt_en.md).
|
3085 |
|
3086 |
-
###
|
3087 |
|
3088 |
```
|
3089 |
baseprompt target1 target2 BREAK
|
@@ -3105,14 +3084,14 @@ is also effective.
|
|
3105 |
|
3106 |
In this example, masks are calculated for shirt, tie, skirt, and color prompts are specified only for those regions.
|
3107 |
|
3108 |
-
```
|
3109 |
rp_args = {
|
3110 |
-
"mode":"prompt-ex",
|
3111 |
-
"save_mask":True,
|
3112 |
"th": "0.4,0.6,0.6",
|
3113 |
}
|
3114 |
|
3115 |
-
prompt ="""
|
3116 |
a girl in street with shirt, tie, skirt BREAK
|
3117 |
red, shirt BREAK
|
3118 |
green, tie BREAK
|
@@ -3122,7 +3101,7 @@ blue , skirt
|
|
3122 |
|
3123 |

|
3124 |
|
3125 |
-
###
|
3126 |
|
3127 |
The threshold used to determine the mask created by the prompt. This can be set as many times as there are masks, as the range varies widely depending on the target prompt. If multiple regions are used, enter them separated by commas. For example, hair tends to be ambiguous and requires a small value, while face tends to be large and requires a small value. These should be ordered by BREAK.
|
3128 |
|
@@ -3141,7 +3120,7 @@ The difference is that in Prompt, duplicate regions are added, whereas in Prompt
|
|
3141 |
|
3142 |
### Accuracy
|
3143 |
|
3144 |
-
In the case of a
|
3145 |
|
3146 |
```
|
3147 |
girl hair twintail frills,ribbons, dress, face BREAK
|
@@ -3154,7 +3133,7 @@ When an image is generated, the generated mask is displayed. It is generated at
|
|
3154 |
|
3155 |
### Use common prompt
|
3156 |
|
3157 |
-
You can attach the prompt up to ADDCOMM to all prompts by separating it first with ADDCOMM. This is useful when you want to include elements common to all regions. For example, when generating pictures of three people with different appearances, it's necessary to include the instruction of 'three people' in all regions. It's also useful when inserting quality tags and other things."For example, if you write as follows:
|
3158 |
|
3159 |
```
|
3160 |
best quality, 3persons in garden, ADDCOMM
|
@@ -3177,24 +3156,24 @@ Negative prompts are equally effective across all regions, but it is possible to
|
|
3177 |
|
3178 |
### Parameters
|
3179 |
|
3180 |
-
To activate Regional Prompter, it is necessary to enter settings in rp_args
|
3181 |
|
3182 |
### Input Parameters
|
3183 |
|
3184 |
Parameters are specified through the `rp_arg`(dictionary type).
|
3185 |
|
3186 |
-
```
|
3187 |
rp_args = {
|
3188 |
"mode":"rows",
|
3189 |
"div": "1;1;1"
|
3190 |
}
|
3191 |
|
3192 |
-
pipe(prompt
|
3193 |
```
|
3194 |
|
3195 |
### Required Parameters
|
3196 |
|
3197 |
-
- `mode`: Specifies the method for defining regions. Choose from `Cols`, `Rows`, `Prompt
|
3198 |
- `divide`: Used in `Cols` and `Rows` modes. Details on how to specify this are provided under the respective `Cols` and `Rows` sections.
|
3199 |
- `th`: Used in `Prompt` mode. The method of specification is detailed under the `Prompt` section.
|
3200 |
|
@@ -3208,7 +3187,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3208 |
|
3209 |
- Reference paper
|
3210 |
|
3211 |
-
```
|
3212 |
@article{chung2022diffusion,
|
3213 |
title={Diffusion posterior sampling for general noisy inverse problems},
|
3214 |
author={Chung, Hyungjin and Kim, Jeongsol and Mccann, Michael T and Klasky, Marc L and Ye, Jong Chul},
|
@@ -3220,7 +3199,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3220 |
- This pipeline allows zero-shot conditional sampling from the posterior distribution $p(x|y)$, given observation on $y$, unconditional generative model $p(x)$ and differentiable operator $y=f(x)$.
|
3221 |
|
3222 |
- For example, $f(.)$ can be downsample operator, then $y$ is a downsampled image, and the pipeline becomes a super-resolution pipeline.
|
3223 |
-
- To use this pipeline, you need to know your operator $f(.)$ and corrupted image $y$, and pass them during the call. For example, as in the main function of dps_pipeline.py
|
3224 |
|
3225 |
```python
|
3226 |
import torch.nn.functional as F
|
@@ -3250,7 +3229,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3250 |
def weights_init(self):
|
3251 |
if self.blur_type == "gaussian":
|
3252 |
n = np.zeros((self.kernel_size, self.kernel_size))
|
3253 |
-
n[self.kernel_size // 2,self.kernel_size // 2] = 1
|
3254 |
k = scipy.ndimage.gaussian_filter(n, sigma=self.std)
|
3255 |
k = torch.from_numpy(k)
|
3256 |
self.k = k
|
@@ -3280,7 +3259,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3280 |
self.conv.update_weights(self.kernel.type(torch.float32))
|
3281 |
|
3282 |
for param in self.parameters():
|
3283 |
-
param.requires_grad=False
|
3284 |
|
3285 |
def forward(self, data, **kwargs):
|
3286 |
return self.conv(data)
|
@@ -3317,7 +3296,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3317 |
- 
|
3318 |
- Gaussian blurred image:
|
3319 |
- 
|
3320 |
-
- You can download those
|
3321 |
|
3322 |
- Next, we need to define a loss function used for diffusion posterior sample. For most of the cases, the RMSE is fine:
|
3323 |
|
@@ -3326,7 +3305,7 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3326 |
return torch.sqrt(torch.sum((yhat-y)**2))
|
3327 |
```
|
3328 |
|
3329 |
-
- And next, as any other diffusion models, we need the score estimator and scheduler. As we are working with $256x256$ face images, we use
|
3330 |
|
3331 |
```python
|
3332 |
# set up scheduler
|
@@ -3343,20 +3322,20 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
|
|
3343 |
# finally, the pipeline
|
3344 |
dpspipe = DPSPipeline(model, scheduler)
|
3345 |
image = dpspipe(
|
3346 |
-
measurement
|
3347 |
-
operator
|
3348 |
-
loss_fn
|
3349 |
-
zeta
|
3350 |
).images[0]
|
3351 |
image.save("dps_generated_image.png")
|
3352 |
```
|
3353 |
|
3354 |
-
- The zeta is a hyperparameter that is in range of $[0,1]$. It
|
3355 |
- Reconstructed image:
|
3356 |
- 
|
3357 |
|
3358 |
- The reconstruction is perceptually similar to the source image, but different in details.
|
3359 |
-
- In dps_pipeline.py
|
3360 |
- Downsampled image:
|
3361 |
- 
|
3362 |
- Reconstructed image:
|
@@ -3368,9 +3347,8 @@ This pipeline combines AnimateDiff and ControlNet. Enjoy precise motion control
|
|
3368 |
|
3369 |
```py
|
3370 |
import torch
|
3371 |
-
from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter
|
3372 |
-
from diffusers.
|
3373 |
-
from diffusers.schedulers import DPMSolverMultistepScheduler
|
3374 |
from PIL import Image
|
3375 |
|
3376 |
motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
|
@@ -3385,7 +3363,8 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
3385 |
controlnet=controlnet,
|
3386 |
vae=vae,
|
3387 |
custom_pipeline="pipeline_animatediff_controlnet",
|
3388 |
-
|
|
|
3389 |
pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
|
3390 |
model_id, subfolder="scheduler", beta_schedule="linear", clip_sample=False, timestep_spacing="linspace", steps_offset=1
|
3391 |
)
|
@@ -3406,7 +3385,6 @@ result = pipe(
|
|
3406 |
num_inference_steps=20,
|
3407 |
).frames[0]
|
3408 |
|
3409 |
-
from diffusers.utils import export_to_gif
|
3410 |
export_to_gif(result.frames[0], "result.gif")
|
3411 |
```
|
3412 |
|
@@ -3431,9 +3409,8 @@ You can also use multiple controlnets at once!
|
|
3431 |
|
3432 |
```python
|
3433 |
import torch
|
3434 |
-
from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter
|
3435 |
-
from diffusers.
|
3436 |
-
from diffusers.schedulers import DPMSolverMultistepScheduler
|
3437 |
from PIL import Image
|
3438 |
|
3439 |
motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
|
@@ -3449,7 +3426,8 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
3449 |
controlnet=[controlnet1, controlnet2],
|
3450 |
vae=vae,
|
3451 |
custom_pipeline="pipeline_animatediff_controlnet",
|
3452 |
-
|
|
|
3453 |
pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
|
3454 |
model_id, subfolder="scheduler", clip_sample=False, timestep_spacing="linspace", steps_offset=1, beta_schedule="linear",
|
3455 |
)
|
@@ -3496,7 +3474,6 @@ result = pipe(
|
|
3496 |
num_inference_steps=20,
|
3497 |
)
|
3498 |
|
3499 |
-
from diffusers.utils import export_to_gif
|
3500 |
export_to_gif(result.frames[0], "result.gif")
|
3501 |
```
|
3502 |
|
@@ -3625,7 +3602,6 @@ pipe.train_lora(prompt, image)
|
|
3625 |
output = pipe(prompt, image, mask_image, source_points, target_points)
|
3626 |
output_image = PIL.Image.fromarray(output)
|
3627 |
output_image.save("./output.png")
|
3628 |
-
|
3629 |
```
|
3630 |
|
3631 |
### Instaflow Pipeline
|
@@ -3674,19 +3650,19 @@ This pipeline provides null-text inversion for editing real images. It enables n
|
|
3674 |
|
3675 |
- Reference paper
|
3676 |
|
3677 |
-
|
3678 |
-
|
3679 |
-
|
3680 |
-
|
3681 |
-
|
|
|
3682 |
```}
|
3683 |
|
3684 |
```py
|
3685 |
-
from diffusers
|
3686 |
from examples.community.pipeline_null_text_inversion import NullTextPipeline
|
3687 |
import torch
|
3688 |
|
3689 |
-
# Load the pipeline
|
3690 |
device = "cuda"
|
3691 |
# Provide invert_prompt and the image for null-text optimization.
|
3692 |
invert_prompt = "A lying cat"
|
@@ -3698,13 +3674,13 @@ prompt = "A lying cat"
|
|
3698 |
# or different if editing.
|
3699 |
prompt = "A lying dog"
|
3700 |
|
3701 |
-
#Float32 is essential to a well optimization
|
3702 |
model_path = "runwayml/stable-diffusion-v1-5"
|
3703 |
scheduler = DDIMScheduler(num_train_timesteps=1000, beta_start=0.00085, beta_end=0.0120, beta_schedule="scaled_linear")
|
3704 |
-
pipeline = NullTextPipeline.from_pretrained(model_path, scheduler
|
3705 |
|
3706 |
-
#Saves the inverted_latent to save time
|
3707 |
-
inverted_latent, uncond = pipeline.invert(input_image, invert_prompt, num_inner_steps=10, early_stop_epsilon=
|
3708 |
pipeline(prompt, uncond, inverted_latent, guidance_scale=7.5, num_inference_steps=steps).images[0].save(input_image+".output.jpg")
|
3709 |
```
|
3710 |
|
@@ -3761,7 +3737,7 @@ for frame in frames:
|
|
3761 |
controlnet = ControlNetModel.from_pretrained(
|
3762 |
"lllyasviel/sd-controlnet-canny").to('cuda')
|
3763 |
|
3764 |
-
# You can use any
|
3765 |
pipe = DiffusionPipeline.from_pretrained(
|
3766 |
"runwayml/stable-diffusion-v1-5", controlnet=controlnet, custom_pipeline='rerender_a_video').to('cuda')
|
3767 |
|
@@ -3803,7 +3779,7 @@ This pipeline is the implementation of [Style Aligned Image Generation via Share
|
|
3803 |
from typing import List
|
3804 |
|
3805 |
import torch
|
3806 |
-
from diffusers
|
3807 |
from PIL import Image
|
3808 |
|
3809 |
model_id = "a-r-r-o-w/dreamshaper-xl-turbo"
|
@@ -3872,7 +3848,7 @@ output = pipe(
|
|
3872 |
image=image,
|
3873 |
prompt="A snail moving on the ground",
|
3874 |
strength=0.8,
|
3875 |
-
latent_interpolation_method="slerp",
|
3876 |
)
|
3877 |
frames = output.frames[0]
|
3878 |
export_to_gif(frames, "animation.gif")
|
@@ -3882,11 +3858,10 @@ export_to_gif(frames, "animation.gif")
|
|
3882 |
|
3883 |
IP Adapter FaceID is an experimental IP Adapter model that uses image embeddings generated by `insightface`, so no image encoder needs to be loaded.
|
3884 |
You need to install `insightface` and all its requirements to use this model.
|
3885 |
-
You must pass the image embedding tensor as `image_embeds` to the
|
3886 |
You can find more results [here](https://github.com/huggingface/diffusers/pull/6276).
|
3887 |
|
3888 |
```py
|
3889 |
-
import diffusers
|
3890 |
import torch
|
3891 |
from diffusers.utils import load_image
|
3892 |
import cv2
|
@@ -3916,7 +3891,7 @@ pipeline.load_ip_adapter_face_id("h94/IP-Adapter-FaceID", "ip-adapter-faceid_sd1
|
|
3916 |
pipeline.to("cuda")
|
3917 |
|
3918 |
generator = torch.Generator(device="cpu").manual_seed(42)
|
3919 |
-
num_images=2
|
3920 |
|
3921 |
image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png")
|
3922 |
|
@@ -3939,13 +3914,13 @@ for i in range(num_images):
|
|
3939 |
|
3940 |
### InstantID Pipeline
|
3941 |
|
3942 |
-
InstantID is a new state-of-the-art tuning-free method to achieve ID-Preserving generation with only single image, supporting various downstream tasks. For any
|
3943 |
|
3944 |
```py
|
3945 |
-
# !pip install opencv-python transformers accelerate insightface
|
3946 |
import diffusers
|
3947 |
from diffusers.utils import load_image
|
3948 |
-
from diffusers
|
3949 |
|
3950 |
import cv2
|
3951 |
import torch
|
@@ -3963,12 +3938,13 @@ app.prepare(ctx_id=0, det_size=(640, 640))
|
|
3963 |
# prepare models under ./checkpoints
|
3964 |
# https://huggingface.co/InstantX/InstantID
|
3965 |
from huggingface_hub import hf_hub_download
|
|
|
3966 |
hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/config.json", local_dir="./checkpoints")
|
3967 |
hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/diffusion_pytorch_model.safetensors", local_dir="./checkpoints")
|
3968 |
hf_hub_download(repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="./checkpoints")
|
3969 |
|
3970 |
-
face_adapter =
|
3971 |
-
controlnet_path =
|
3972 |
|
3973 |
# load IdentityNet
|
3974 |
controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
|
@@ -3979,7 +3955,7 @@ pipe = StableDiffusionXLInstantIDPipeline.from_pretrained(
|
|
3979 |
controlnet=controlnet,
|
3980 |
torch_dtype=torch.float16
|
3981 |
)
|
3982 |
-
pipe.cuda
|
3983 |
|
3984 |
# load adapter
|
3985 |
pipe.load_ip_adapter_instantid(face_adapter)
|
@@ -4046,8 +4022,9 @@ import cv2
|
|
4046 |
import torch
|
4047 |
import numpy as np
|
4048 |
|
4049 |
-
from diffusers import ControlNetModel,DDIMScheduler, DiffusionPipeline
|
4050 |
import sys
|
|
|
4051 |
gmflow_dir = "/path/to/gmflow"
|
4052 |
sys.path.insert(0, gmflow_dir)
|
4053 |
|
@@ -4075,7 +4052,7 @@ def video_to_frame(video_path: str, interval: int):
|
|
4075 |
input_video_path = 'https://github.com/williamyang1991/FRESCO/raw/main/data/car-turn.mp4'
|
4076 |
output_video_path = 'car.gif'
|
4077 |
|
4078 |
-
# You can use any
|
4079 |
model_path = 'SG161222/Realistic_Vision_V2.0'
|
4080 |
|
4081 |
prompt = 'a red car turns in the winter'
|
@@ -4120,14 +4097,13 @@ output_frames = pipe(
|
|
4120 |
|
4121 |
output_frames[0].save(output_video_path, save_all=True,
|
4122 |
append_images=output_frames[1:], duration=100, loop=0)
|
4123 |
-
|
4124 |
```
|
4125 |
|
4126 |
# Perturbed-Attention Guidance
|
4127 |
|
4128 |
[Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)
|
4129 |
|
4130 |
-
This implementation is based on [Diffusers](https://huggingface.co/docs/diffusers/index). StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).
|
4131 |
|
4132 |
## Example Usage
|
4133 |
|
@@ -4147,14 +4123,14 @@ pipe = StableDiffusionPipeline.from_pretrained(
|
|
4147 |
torch_dtype=torch.float16
|
4148 |
)
|
4149 |
|
4150 |
-
device="cuda"
|
4151 |
pipe = pipe.to(device)
|
4152 |
|
4153 |
pag_scale = 5.0
|
4154 |
pag_applied_layers_index = ['m0']
|
4155 |
|
4156 |
batch_size = 4
|
4157 |
-
seed=10
|
4158 |
|
4159 |
base_dir = "./results/"
|
4160 |
grid_dir = base_dir + "/pag" + str(pag_scale) + "/"
|
@@ -4164,7 +4140,7 @@ if not os.path.exists(grid_dir):
|
|
4164 |
|
4165 |
set_seed(seed)
|
4166 |
|
4167 |
-
latent_input = randn_tensor(shape=(batch_size,4,64,64),generator=None, device=device, dtype=torch.float16)
|
4168 |
|
4169 |
output_baseline = pipe(
|
4170 |
"",
|
@@ -4196,6 +4172,6 @@ grid_image.save(grid_dir + "sample.png")
|
|
4196 |
|
4197 |
## PAG Parameters
|
4198 |
|
4199 |
-
pag_scale :
|
4200 |
|
4201 |
-
pag_applied_layers_index : index of the layer to apply perturbation (ex: ['m0'])
|
|
|
27 |
| Multilingual Stable Diffusion | Stable Diffusion Pipeline that supports prompts in 50 different languages. | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline) | - | [Juan Carlos Piñeros](https://github.com/juancopi81) |
|
28 |
| GlueGen Stable Diffusion | Stable Diffusion Pipeline that supports prompts in different languages using GlueGen adapter. | [GlueGen Stable Diffusion](#gluegen-stable-diffusion-pipeline) | - | [Phạm Hồng Vinh](https://github.com/rootonchair) |
|
29 |
| Image to Image Inpainting Stable Diffusion | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | - | [Alex McKinney](https://github.com/vvvm23) |
|
30 |
+
| Text Based Inpainting Stable Diffusion | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting | [Text Based Inpainting Stable Diffusion](#text-based-inpainting-stable-diffusion) | - | [Dhruv Karan](https://github.com/unography) |
|
31 |
| Bit Diffusion | Diffusion on discrete data | [Bit Diffusion](#bit-diffusion) | - | [Stuti R.](https://github.com/kingstut) |
|
32 |
| K-Diffusion Stable Diffusion | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py) | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
|
33 |
| Checkpoint Merger Pipeline | Diffusion Pipeline that enables merging of saved model checkpoints | [Checkpoint Merger Pipeline](#checkpoint-merger-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
|
|
|
40 |
| CLIP Guided Img2Img Stable Diffusion Pipeline | Doing CLIP guidance for image to image generation with Stable Diffusion | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion) | - | [Nipun Jindal](https://github.com/nipunjindal/) |
|
41 |
| TensorRT Stable Diffusion Text to Image Pipeline | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
|
42 |
| EDICT Image Editing Pipeline | Diffusion pipeline for text-guided image editing | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline) | - | [Joqsan Azocar](https://github.com/Joqsan) |
|
43 |
+
| Stable Diffusion RePaint | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.09865) for inpainting. | [Stable Diffusion RePaint](#stable-diffusion-repaint ) | - | [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
|
44 |
| TensorRT Stable Diffusion Image to Image Pipeline | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
|
45 |
| Stable Diffusion IPEX Pipeline | Accelerate Stable Diffusion inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion on IPEX](#stable-diffusion-on-ipex) | - | [Yingjie Han](https://github.com/yingjie-han/) |
|
46 |
| CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |
|
|
|
192 |
init_image = load_image("https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/samples/images/2.jpg")
|
193 |
mask_image = load_image("https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/samples/masks/2.png")
|
194 |
|
195 |
+
image = pipe(prompt, init_image, mask_image, use_rasg=True, use_painta=True, generator=torch.manual_seed(12345)).images[0]
|
196 |
|
197 |
make_image_grid([init_image, mask_image, image], rows=1, cols=3)
|
|
|
198 |
```
|
199 |
|
200 |
### Marigold Depth Estimation
|
|
|
222 |
|
223 |
# (New) LCM version (faster speed)
|
224 |
pipe = DiffusionPipeline.from_pretrained(
|
225 |
+
"prs-eth/marigold-depth-lcm-v1-0",
|
226 |
custom_pipeline="marigold_depth_estimation"
|
227 |
# torch_dtype=torch.float16, # (optional) Run with half-precision (16-bit float).
|
228 |
# variant="fp16", # (optional) Use with `torch_dtype=torch.float16`, to directly load fp16 checkpoint
|
|
|
365 |
custom_pipeline="clip_guided_stable_diffusion",
|
366 |
clip_model=clip_model,
|
367 |
feature_extractor=feature_extractor,
|
|
|
368 |
torch_dtype=torch.float16,
|
369 |
)
|
370 |
guided_pipeline.enable_attention_slicing()
|
|
|
392 |
```
|
393 |
|
394 |
The `images` list contains a list of PIL images that can be saved locally or displayed directly in a google colab.
|
395 |
+
Generated images tend to be of higher quality than natively using stable diffusion. E.g. the above script generates the following images:
|
396 |
|
397 |
.
|
398 |
|
|
|
466 |
|
467 |
|
468 |
### Text-to-Image
|
|
|
469 |
images = pipe.text2img("An astronaut riding a horse").images
|
470 |
|
471 |
### Image-to-Image
|
|
|
472 |
init_image = download_image("https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg")
|
473 |
|
474 |
prompt = "A fantasy landscape, trending on artstation"
|
|
|
476 |
images = pipe.img2img(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images
|
477 |
|
478 |
### Inpainting
|
|
|
479 |
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
|
480 |
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
|
481 |
init_image = download_image(img_url).resize((512, 512))
|
|
|
492 |
Features of this custom pipeline:
|
493 |
|
494 |
- Input a prompt without the 77 token length limit.
|
495 |
+
- Includes tx2img, img2img, and inpainting pipelines.
|
496 |
- Emphasize/weigh part of your prompt with parentheses as so: `a baby deer with (big eyes)`
|
497 |
- De-emphasize part of your prompt as so: `a [baby] deer with big eyes`
|
498 |
- Precisely weigh part of your prompt as so: `a baby deer with (big eyes:1.3)`
|
|
|
506 |
|
507 |
You can run this custom pipeline as so:
|
508 |
|
509 |
+
#### PyTorch
|
510 |
|
511 |
```python
|
512 |
from diffusers import DiffusionPipeline
|
|
|
515 |
pipe = DiffusionPipeline.from_pretrained(
|
516 |
'hakurei/waifu-diffusion',
|
517 |
custom_pipeline="lpw_stable_diffusion",
|
|
|
518 |
torch_dtype=torch.float16
|
519 |
)
|
520 |
+
pipe = pipe.to("cuda")
|
521 |
|
522 |
prompt = "best_quality (1girl:1.3) bow bride brown_hair closed_mouth frilled_bow frilled_hair_tubes frills (full_body:1.3) fox_ear hair_bow hair_tubes happy hood japanese_clothes kimono long_sleeves red_bow smile solo tabi uchikake white_kimono wide_sleeves cherry_blossoms"
|
523 |
neg_prompt = "lowres, bad_anatomy, error_body, error_hair, error_arm, error_hands, bad_hands, error_fingers, bad_fingers, missing_fingers, error_legs, bad_legs, multiple_legs, missing_legs, error_lighting, error_shadow, error_reflection, text, error, extra_digit, fewer_digits, cropped, worst_quality, low_quality, normal_quality, jpeg_artifacts, signature, watermark, username, blurry"
|
524 |
|
525 |
+
pipe.text2img(prompt, negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
|
|
|
526 |
```
|
527 |
|
528 |
#### onnxruntime
|
|
|
541 |
prompt = "a photo of an astronaut riding a horse on mars, best quality"
|
542 |
neg_prompt = "lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
|
543 |
|
544 |
+
pipe.text2img(prompt, negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
|
|
|
545 |
```
|
546 |
|
547 |
+
If you see `Token indices sequence length is longer than the specified maximum sequence length for this model ( *** > 77 ) . Running this sequence through the model will result in indexing errors`. Do not worry, it is normal.
|
548 |
|
549 |
### Speech to Image
|
550 |
|
|
|
579 |
custom_pipeline="speech_to_image_diffusion",
|
580 |
speech_model=model,
|
581 |
speech_processor=processor,
|
|
|
582 |
torch_dtype=torch.float16,
|
583 |
)
|
584 |
|
|
|
638 |
pipe = DiffusionPipeline.from_pretrained(
|
639 |
"CompVis/stable-diffusion-v1-4",
|
640 |
custom_pipeline="wildcard_stable_diffusion",
|
|
|
641 |
torch_dtype=torch.float16,
|
642 |
)
|
643 |
prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
|
|
|
697 |
images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
|
698 |
grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
|
699 |
tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
|
|
|
700 |
```
|
701 |
|
702 |
### Imagic Stable Diffusion
|
|
|
710 |
import torch
|
711 |
import os
|
712 |
from diffusers import DiffusionPipeline, DDIMScheduler
|
713 |
+
|
714 |
has_cuda = torch.cuda.is_available()
|
715 |
device = torch.device('cpu' if not has_cuda else 'cuda')
|
716 |
pipe = DiffusionPipeline.from_pretrained(
|
717 |
"CompVis/stable-diffusion-v1-4",
|
718 |
+
safety_checker=None,
|
719 |
custom_pipeline="imagic_stable_diffusion",
|
720 |
+
scheduler=DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
|
721 |
).to(device)
|
722 |
generator = torch.Generator("cuda").manual_seed(0)
|
723 |
seed = 0
|
|
|
827 |
|
828 |
### Multilingual Stable Diffusion Pipeline
|
829 |
|
830 |
+
The following code can generate images from texts in different languages using the pre-trained [mBART-50 many-to-one multilingual machine translation model](https://huggingface.co/facebook/mbart-large-50-many-to-one-mmt) and Stable Diffusion.
|
831 |
|
832 |
```python
|
833 |
from PIL import Image
|
|
|
871 |
detection_pipeline=language_detection_pipeline,
|
872 |
translation_model=trans_model,
|
873 |
translation_tokenizer=trans_tokenizer,
|
|
|
874 |
torch_dtype=torch.float16,
|
875 |
)
|
876 |
|
|
|
894 |
|
895 |
### GlueGen Stable Diffusion Pipeline
|
896 |
|
897 |
+
GlueGen is a minimal adapter that allows alignment between any encoder (Text Encoder of different language, Multilingual Roberta, AudioClip) and CLIP text encoder used in standard Stable Diffusion model. This method allows easy language adaptation to available english Stable Diffusion checkpoints without the need of an image captioning dataset as well as long training hours.
|
898 |
|
899 |
+
Make sure you downloaded `gluenet_French_clip_overnorm_over3_noln.ckpt` for French (there are also pre-trained weights for Chinese, Italian, Japanese, Spanish or train your own) at [GlueGen's official repo](https://github.com/salesforce/GlueGen/tree/main).
|
900 |
|
901 |
```python
|
902 |
from PIL import Image
|
|
|
963 |
pipe = DiffusionPipeline.from_pretrained(
|
964 |
"runwayml/stable-diffusion-inpainting",
|
965 |
custom_pipeline="img2img_inpainting",
|
|
|
966 |
torch_dtype=torch.float16
|
967 |
)
|
968 |
pipe = pipe.to("cuda")
|
|
|
1007 |
|
1008 |
### Bit Diffusion
|
1009 |
|
1010 |
+
Based <https://arxiv.org/abs/2208.04202>, this is used for diffusion on discrete data - eg, discrete image data, DNA sequence data. An unconditional discrete image can be generated like this:
|
1011 |
|
1012 |
```python
|
1013 |
from diffusers import DiffusionPipeline
|
1014 |
+
|
1015 |
pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="bit_diffusion")
|
1016 |
image = pipe().images[0]
|
|
|
1017 |
```
|
1018 |
|
1019 |
### Stable Diffusion with K Diffusion
|
|
|
1079 |
|
1080 |
### Checkpoint Merger Pipeline
|
1081 |
|
1082 |
+
Based on the AUTOMATIC1111/webui for checkpoint merging. This is a custom pipeline that merges up to 3 pretrained model checkpoints as long as they are in the HuggingFace model_index.json format.
|
1083 |
|
1084 |
+
The checkpoint merging is currently memory intensive as it modifies the weights of a DiffusionPipeline object in place. Expect at least 13GB RAM usage on Kaggle GPU kernels and
|
1085 |
+
on Colab you might run out of the 12GB memory even while merging two checkpoints.
|
1086 |
|
1087 |
Usage:-
|
1088 |
|
1089 |
```python
|
1090 |
from diffusers import DiffusionPipeline
|
1091 |
|
1092 |
+
# Return a CheckpointMergerPipeline class that allows you to merge checkpoints.
|
1093 |
+
# The checkpoint passed here is ignored. But still pass one of the checkpoints you plan to
|
1094 |
+
# merge for convenience
|
1095 |
pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="checkpoint_merger")
|
1096 |
|
1097 |
+
# There are multiple possible scenarios:
|
1098 |
+
# The pipeline with the merged checkpoints is returned in all the scenarios
|
1099 |
|
1100 |
+
# Compatible checkpoints a.k.a matched model_index.json files. Ignores the meta attributes in model_index.json during comparison.( attrs with _ as prefix )
|
1101 |
+
merged_pipe = pipe.merge(["CompVis/stable-diffusion-v1-4"," CompVis/stable-diffusion-v1-2"], interp="sigmoid", alpha=0.4)
|
1102 |
|
1103 |
+
# Incompatible checkpoints in model_index.json but merge might be possible. Use force=True to ignore model_index.json compatibility
|
1104 |
+
merged_pipe_1 = pipe.merge(["CompVis/stable-diffusion-v1-4", "hakurei/waifu-diffusion"], force=True, interp="sigmoid", alpha=0.4)
|
1105 |
|
1106 |
+
# Three checkpoint merging. Only "add_difference" method actually works on all three checkpoints. Using any other options will ignore the 3rd checkpoint.
|
1107 |
+
merged_pipe_2 = pipe.merge(["CompVis/stable-diffusion-v1-4", "hakurei/waifu-diffusion", "prompthero/openjourney"], force=True, interp="add_difference", alpha=0.4)
|
1108 |
|
1109 |
prompt = "An astronaut riding a horse on Mars"
|
1110 |
|
1111 |
image = merged_pipe(prompt).images[0]
|
|
|
1112 |
```
|
1113 |
|
1114 |
Some examples along with the merge details:
|
|
|
1119 |
|
1120 |
2. "hakurei/waifu-diffusion" + "prompthero/openjourney" ; Inverse Sigmoid interpolation; alpha = 0.8
|
1121 |
|
1122 |
+

|
1123 |
|
1124 |
3. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" + "prompthero/openjourney"; Add Difference interpolation; alpha = 0.5
|
1125 |
|
|
|
1184 |
pipe = DiffusionPipeline.from_pretrained(
|
1185 |
"CompVis/stable-diffusion-v1-4",
|
1186 |
custom_pipeline="magic_mix",
|
1187 |
+
scheduler=DDIMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler"),
|
1188 |
).to('cuda')
|
1189 |
|
1190 |
img = Image.open('phone.jpg')
|
1191 |
mix_img = pipe(
|
1192 |
img,
|
1193 |
+
prompt='bed',
|
1194 |
+
kmin=0.3,
|
1195 |
+
kmax=0.5,
|
1196 |
+
mix_factor=0.5,
|
1197 |
)
|
1198 |
mix_img.save('phone_bed_mix.jpg')
|
1199 |
```
|
|
|
1214 |
|
1215 |
### Stable UnCLIP
|
1216 |
|
1217 |
+
UnCLIPPipeline("kakaobrain/karlo-v1-alpha") provides a prior model that can generate clip image embedding from text.
|
1218 |
+
StableDiffusionImageVariationPipeline("lambdalabs/sd-image-variations-diffusers") provides a decoder model than can generate images from clip image embedding.
|
1219 |
|
1220 |
```python
|
1221 |
import torch
|
|
|
1256 |
print(pipeline.decoder_pipe.__class__)
|
1257 |
# <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline'>
|
1258 |
|
1259 |
+
# this pipeline only uses prior module in "kakaobrain/karlo-v1-alpha"
|
1260 |
# It is used to convert clip text embedding to clip image embedding.
|
1261 |
print(pipeline)
|
1262 |
# StableUnCLIPPipeline {
|
|
|
1316 |
|
1317 |
start_prompt = "A photograph of an adult lion"
|
1318 |
end_prompt = "A photograph of a lion cub"
|
1319 |
+
# For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
|
1320 |
generator = torch.Generator(device=device).manual_seed(42)
|
1321 |
|
1322 |
+
output = pipe(start_prompt, end_prompt, steps=6, generator=generator, enable_sequential_cpu_offload=False)
|
1323 |
|
1324 |
for i,image in enumerate(output.images):
|
1325 |
img.save('result%s.jpg' % i)
|
|
|
1354 |
pipe.to(device)
|
1355 |
|
1356 |
images = [Image.open('./starry_night.jpg'), Image.open('./flowers.jpg')]
|
1357 |
+
# For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
|
1358 |
generator = torch.Generator(device=device).manual_seed(42)
|
1359 |
|
1360 |
+
output = pipe(image=images, steps=6, generator=generator)
|
1361 |
|
1362 |
for i,image in enumerate(output.images):
|
1363 |
image.save('starry_to_flowers_%s.jpg' % i)
|
|
|
1379 |
|
1380 |
### DDIM Noise Comparative Analysis Pipeline
|
1381 |
|
1382 |
+
#### **Research question: What visual concepts do the diffusion models learn from each noise level during training?**
|
1383 |
|
1384 |
The [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227) paper proposed an approach to answer the above question, which is their second contribution.
|
1385 |
The approach consists of the following steps:
|
|
|
1396 |
from PIL import Image
|
1397 |
import numpy as np
|
1398 |
|
1399 |
+
image_path = "path/to/your/image" # images from CelebA-HQ might be better
|
1400 |
image_pil = Image.open(image_path)
|
1401 |
image_name = image_path.split("/")[-1].split(".")[0]
|
1402 |
|
|
|
1435 |
from diffusers import DiffusionPipeline
|
1436 |
from PIL import Image
|
1437 |
from transformers import CLIPFeatureExtractor, CLIPModel
|
1438 |
+
|
1439 |
feature_extractor = CLIPFeatureExtractor.from_pretrained(
|
1440 |
"laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
|
1441 |
)
|
|
|
1610 |
import torch
|
1611 |
from io import BytesIO
|
1612 |
from diffusers import StableDiffusionPipeline, RePaintScheduler
|
1613 |
+
|
1614 |
def download_image(url):
|
1615 |
response = requests.get(url)
|
1616 |
return PIL.Image.open(BytesIO(response.content)).convert("RGB")
|
|
|
1668 |
```
|
1669 |
|
1670 |
### Stable Diffusion BoxDiff
|
1671 |
+
BoxDiff is a training-free method for controlled generation with bounding box coordinates. It should work with any Stable Diffusion model. Below shows an example with `stable-diffusion-2-1-base`.
|
1672 |
```py
|
1673 |
import torch
|
1674 |
from PIL import Image, ImageDraw
|
|
|
1828 |
|
1829 |
### Stable Diffusion on IPEX
|
1830 |
|
1831 |
+
This diffusion pipeline aims to accelerate the inference of Stable-Diffusion on Intel Xeon CPUs with BF16/FP32 precision using [IPEX](https://github.com/intel/intel-extension-for-pytorch).
|
1832 |
|
1833 |
To use this pipeline, you need to:
|
1834 |
|
1835 |
1. Install [IPEX](https://github.com/intel/intel-extension-for-pytorch)
|
1836 |
|
1837 |
+
**Note:** For each PyTorch release, there is a corresponding release of the IPEX. Here is the mapping relationship. It is recommended to install PyTorch/IPEX2.0 to get the best performance.
|
1838 |
|
1839 |
|PyTorch Version|IPEX Version|
|
1840 |
|--|--|
|
|
|
1853 |
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
|
1854 |
```
|
1855 |
|
1856 |
+
2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX acceleration. Supported inference datatypes are Float32 and BFloat16.
|
1857 |
|
1858 |
**Note:** The setting of generated image height/width for `prepare_for_ipex()` should be same as the setting of pipeline inference.
|
1859 |
|
1860 |
```python
|
1861 |
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_ipex")
|
1862 |
# For Float32
|
1863 |
+
pipe.prepare_for_ipex(prompt, dtype=torch.float32, height=512, width=512) # value of image height/width should be consistent with the pipeline inference
|
1864 |
# For BFloat16
|
1865 |
+
pipe.prepare_for_ipex(prompt, dtype=torch.bfloat16, height=512, width=512) # value of image height/width should be consistent with the pipeline inference
|
1866 |
```
|
1867 |
|
1868 |
Then you can use the ipex pipeline in a similar way to the default stable diffusion pipeline.
|
1869 |
|
1870 |
```python
|
1871 |
# For Float32
|
1872 |
+
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] # value of image height/width should be consistent with 'prepare_for_ipex()'
|
1873 |
# For BFloat16
|
1874 |
with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
1875 |
+
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] # value of image height/width should be consistent with 'prepare_for_ipex()'
|
1876 |
```
|
1877 |
|
1878 |
The following code compares the performance of the original stable diffusion pipeline with the ipex-optimized pipeline.
|
|
|
1890 |
# warmup
|
1891 |
for _ in range(2):
|
1892 |
images = pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512).images
|
1893 |
+
# time evaluation
|
1894 |
start = time.time()
|
1895 |
for _ in range(nb_pass):
|
1896 |
pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512)
|
|
|
1911 |
latency = elapsed_time(pipe)
|
1912 |
print("Latency of StableDiffusionIPEXPipeline--bf16", latency)
|
1913 |
latency = elapsed_time(pipe2)
|
1914 |
+
print("Latency of StableDiffusionPipeline--bf16", latency)
|
1915 |
|
1916 |
############## fp32 inference performance ###############
|
1917 |
|
|
|
1926 |
latency = elapsed_time(pipe3)
|
1927 |
print("Latency of StableDiffusionIPEXPipeline--fp32", latency)
|
1928 |
latency = elapsed_time(pipe4)
|
1929 |
+
print("Latency of StableDiffusionPipeline--fp32", latency)
|
|
|
1930 |
```
|
1931 |
|
1932 |
### Stable Diffusion XL on IPEX
|
1933 |
|
1934 |
+
This diffusion pipeline aims to accelerate the inference of Stable-Diffusion XL on Intel Xeon CPUs with BF16/FP32 precision using [IPEX](https://github.com/intel/intel-extension-for-pytorch).
|
1935 |
|
1936 |
To use this pipeline, you need to:
|
1937 |
|
|
|
1956 |
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
|
1957 |
```
|
1958 |
|
1959 |
+
2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX acceleration. Supported inference datatypes are Float32 and BFloat16.
|
1960 |
|
1961 |
**Note:** The values of `height` and `width` used during preparation with `prepare_for_ipex()` should be the same when running inference with the prepared pipeline.
|
1962 |
|
|
|
1999 |
# warmup
|
2000 |
for _ in range(2):
|
2001 |
images = pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512, guidance_scale=0.0).images
|
2002 |
+
# time evaluation
|
2003 |
start = time.time()
|
2004 |
for _ in range(nb_pass):
|
2005 |
pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512, guidance_scale=0.0)
|
|
|
2035 |
latency = elapsed_time(pipe3, num_inference_steps=steps)
|
2036 |
print("Latency of StableDiffusionXLPipelineIpex--fp32", latency, "s for total", steps, "steps")
|
2037 |
latency = elapsed_time(pipe4, num_inference_steps=steps)
|
2038 |
+
print("Latency of StableDiffusionXLPipeline--fp32", latency, "s for total", steps, "steps")
|
|
|
2039 |
```
|
2040 |
|
2041 |
### CLIP Guided Images Mixing With Stable Diffusion
|
|
|
2048 |
|
2049 |
### Stable Diffusion XL Long Weighted Prompt Pipeline
|
2050 |
|
2051 |
+
This SDXL pipeline supports unlimited length prompt and negative prompt, compatible with A1111 prompt weighted style.
|
2052 |
|
2053 |
You can provide both `prompt` and `prompt_2`. If only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline.
|
2054 |
|
|
|
2076 |
t2i_images = pipe(
|
2077 |
prompt=prompt,
|
2078 |
negative_prompt=neg_prompt,
|
2079 |
+
).images # alternatively, you can call the .text2img() function
|
2080 |
|
2081 |
# img2img
|
2082 |
+
input_image = load_image("/path/to/local/image.png") # or URL to your input image
|
2083 |
i2i_images = pipe.img2img(
|
2084 |
prompt=prompt,
|
2085 |
negative_prompt=neg_prompt,
|
2086 |
image=input_image,
|
2087 |
+
strength=0.8, # higher strength will result in more variation compared to original image
|
2088 |
).images
|
2089 |
|
2090 |
# inpaint
|
2091 |
+
input_mask = load_image("/path/to/local/mask.png") # or URL to your input inpainting mask
|
2092 |
inpaint_images = pipe.inpaint(
|
2093 |
prompt="photo of a cute (black) cat running on the grass" * 20,
|
2094 |
negative_prompt=neg_prompt,
|
2095 |
image=input_image,
|
2096 |
mask=input_mask,
|
2097 |
+
strength=0.6, # higher strength will result in more variation compared to original image
|
2098 |
).images
|
2099 |
|
2100 |
pipe.to("cpu")
|
2101 |
torch.cuda.empty_cache()
|
2102 |
|
2103 |
+
from IPython.display import display # assuming you are using this code in a notebook
|
2104 |
display(t2i_images[0])
|
2105 |
display(i2i_images[0])
|
2106 |
display(inpaint_images[0])
|
|
|
2140 |
coca_model.dtype = torch.float16
|
2141 |
coca_transform = open_clip.image_transform(
|
2142 |
coca_model.visual.image_size,
|
2143 |
+
is_train=False,
|
2144 |
+
mean=getattr(coca_model.visual, 'image_mean', None),
|
2145 |
+
std=getattr(coca_model.visual, 'image_std', None),
|
2146 |
)
|
2147 |
coca_tokenizer = SimpleTokenizer()
|
2148 |
|
|
|
2194 |
```python
|
2195 |
from diffusers import LMSDiscreteScheduler, DiffusionPipeline
|
2196 |
|
2197 |
+
# Create scheduler and model (similar to StableDiffusionPipeline)
|
2198 |
scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
|
2199 |
pipeline = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler, custom_pipeline="mixture_tiling")
|
2200 |
pipeline.to("cuda")
|
|
|
2235 |
# Use the PNDMScheduler scheduler here instead
|
2236 |
scheduler = PNDMScheduler.from_pretrained("stabilityai/stable-diffusion-2-inpainting", subfolder="scheduler")
|
2237 |
|
|
|
2238 |
pipe = StableDiffusionInpaintPipeline.from_pretrained("stabilityai/stable-diffusion-2-inpainting",
|
2239 |
custom_pipeline="stable_diffusion_tensorrt_inpaint",
|
2240 |
variant='fp16',
|
|
|
2273 |
# Load and preprocess guide image
|
2274 |
iic_image = preprocess_image(Image.open("input_image.png").convert("RGB"))
|
2275 |
|
2276 |
+
# Create scheduler and model (similar to StableDiffusionPipeline)
|
2277 |
scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
|
2278 |
pipeline = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler).to("cuda:0", custom_pipeline="mixture_canvas")
|
2279 |
pipeline.to("cuda")
|
|
|
2284 |
canvas_width=352,
|
2285 |
regions=[
|
2286 |
Text2ImageRegion(0, 800, 0, 352, guidance_scale=8,
|
2287 |
+
prompt=f"best quality, masterpiece, WLOP, sakimichan, art contest winner on pixiv, 8K, intricate details, wet effects, rain drops, ethereal, mysterious, futuristic, UHD, HDR, cinematic lighting, in a beautiful forest, rainy day, award winning, trending on artstation, beautiful confident cheerful young woman, wearing a futuristic sleeveless dress, ultra beautiful detailed eyes, hyper-detailed face, complex, perfect, model, textured, chiaroscuro, professional make-up, realistic, figure in frame, "),
|
2288 |
Image2ImageRegion(352-800, 352, 0, 352, reference_image=iic_image, strength=1.0),
|
2289 |
],
|
2290 |
num_inference_steps=100,
|
|
|
2303 |
The following code shows how to use the IADB pipeline to generate images using a pretrained celebahq-256 model.
|
2304 |
|
2305 |
```python
|
|
|
2306 |
pipeline_iadb = DiffusionPipeline.from_pretrained("thomasc4/iadb-celebahq-256", custom_pipeline='iadb')
|
2307 |
|
2308 |
pipeline_iadb = pipeline_iadb.to('cuda')
|
2309 |
|
2310 |
+
output = pipeline_iadb(batch_size=4, num_inference_steps=128)
|
2311 |
for i in range(len(output[0])):
|
2312 |
plt.imshow(output[0][i])
|
2313 |
plt.show()
|
|
|
2314 |
```
|
2315 |
|
2316 |
Sampling with the IADB formulation is easy, and can be done in a few lines (the pipeline already implements it):
|
2317 |
|
2318 |
```python
|
|
|
2319 |
def sample_iadb(model, x0, nb_step):
|
2320 |
x_alpha = x0
|
2321 |
for t in range(nb_step):
|
|
|
2326 |
x_alpha = x_alpha + (alpha_next-alpha)*d
|
2327 |
|
2328 |
return x_alpha
|
|
|
2329 |
```
|
2330 |
|
2331 |
The training loop is also straightforward:
|
2332 |
|
2333 |
```python
|
|
|
2334 |
# Training loop
|
2335 |
while True:
|
2336 |
x0 = sample_noise()
|
|
|
2361 |
from pipeline_zero1to3 import Zero1to3StableDiffusionPipeline
|
2362 |
from diffusers.utils import load_image
|
2363 |
|
2364 |
+
model_id = "kxic/zero123-165000" # zero123-105000, zero123-165000, zero123-xl
|
2365 |
|
2366 |
pipe = Zero1to3StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
|
2367 |
|
|
|
2382 |
# H, W = (256, 256) # H, W = (512, 512) # zero123 training is 256,256
|
2383 |
|
2384 |
# for batch input
|
2385 |
+
input_image1 = load_image("./demo/4_blackarm.png") # load_image("https://cvlab-zero123-live.hf.space/file=/home/user/app/configs/4_blackarm.png")
|
2386 |
+
input_image2 = load_image("./demo/8_motor.png") # load_image("https://cvlab-zero123-live.hf.space/file=/home/user/app/configs/8_motor.png")
|
2387 |
+
input_image3 = load_image("./demo/7_london.png") # load_image("https://cvlab-zero123-live.hf.space/file=/home/user/app/configs/7_london.png")
|
2388 |
input_images = [input_image1, input_image2, input_image3]
|
2389 |
query_poses = [query_pose1, query_pose2, query_pose3]
|
2390 |
|
|
|
2415 |
images = pipe(input_imgs=input_images, prompt_imgs=input_images, poses=query_poses, height=H, width=W,
|
2416 |
guidance_scale=3.0, num_images_per_prompt=num_images_per_prompt, num_inference_steps=50).images
|
2417 |
|
|
|
2418 |
# save imgs
|
2419 |
log_dir = "logs"
|
2420 |
os.makedirs(log_dir, exist_ok=True)
|
|
|
2424 |
for idx in range(num_images_per_prompt):
|
2425 |
images[i].save(os.path.join(log_dir,f"obj{obj}_{idx}.jpg"))
|
2426 |
i += 1
|
|
|
2427 |
```
|
2428 |
|
2429 |
### Stable Diffusion XL Reference
|
2430 |
|
2431 |
+
This pipeline uses the Reference. Refer to the [stable_diffusion_reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-reference).
|
2432 |
|
2433 |
```py
|
2434 |
import torch
|
|
|
2436 |
from diffusers.utils import load_image
|
2437 |
from diffusers import DiffusionPipeline
|
2438 |
from diffusers.schedulers import UniPCMultistepScheduler
|
2439 |
+
|
2440 |
input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
|
2441 |
|
2442 |
# pipe = DiffusionPipeline.from_pretrained(
|
|
|
2509 |
# load the pipeline
|
2510 |
# make sure you're logged in with `huggingface-cli login`
|
2511 |
model_id_or_path = "runwayml/stable-diffusion-v1-5"
|
2512 |
+
# can also be used with dreamlike-art/dreamlike-photoreal-2.0
|
2513 |
pipe = DiffusionPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16, custom_pipeline="pipeline_fabric").to("cuda")
|
2514 |
|
2515 |
# let's specify a prompt
|
|
|
2540 |
image = pipe(
|
2541 |
prompt=prompt,
|
2542 |
negative_prompt=negative_prompt,
|
2543 |
+
liked=liked,
|
2544 |
num_inference_steps=20,
|
2545 |
).images[0]
|
2546 |
|
|
|
2710 |
```py
|
2711 |
prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
|
2712 |
|
2713 |
+
# Can be set to 1~50 steps. LCM supports fast inference even <= 4 steps. Recommend: 1~8 steps.
|
2714 |
num_inference_steps = 4
|
2715 |
|
2716 |
images = pipe(prompt=prompt, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
|
|
|
2742 |
|
2743 |
input_image=Image.open("myimg.png")
|
2744 |
|
2745 |
+
strength = 0.5 # strength =0 (no change) strength=1 (completely overwrite image)
|
2746 |
|
2747 |
+
# Can be set to 1~50 steps. LCM supports fast inference even <= 4 steps. Recommend: 1~8 steps.
|
2748 |
num_inference_steps = 4
|
2749 |
|
2750 |
images = pipe(prompt=prompt, image=input_image, strength=strength, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
|
|
|
2788 |
guidance_scale=8.0,
|
2789 |
embedding_interpolation_type="lerp",
|
2790 |
latent_interpolation_type="slerp",
|
2791 |
+
process_batch_size=4, # Make it higher or lower based on your GPU memory
|
2792 |
generator=torch.Generator(seed),
|
2793 |
)
|
2794 |
|
|
|
2807 |
- [ldm3d-pano](https://huggingface.co/Intel/ldm3d-pano). This checkpoint enables the generation of panoramic images and requires the StableDiffusionLDM3DPipeline pipeline to be used.
|
2808 |
- [ldm3d-sr](https://huggingface.co/Intel/ldm3d-sr). This checkpoint enables the upscaling of RGB and depth images. Can be used in cascade after the original LDM3D pipeline using the StableDiffusionUpscaleLDM3DPipeline pipeline.
|
2809 |
|
2810 |
+
```py
|
2811 |
from PIL import Image
|
2812 |
import os
|
2813 |
import torch
|
|
|
2818 |
pipe_ldm3d = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
|
2819 |
pipe_ldm3d.to("cuda")
|
2820 |
|
2821 |
+
prompt = "A picture of some lemons on a table"
|
2822 |
output = pipe_ldm3d(prompt)
|
2823 |
rgb_image, depth_image = output.rgb, output.depth
|
2824 |
+
rgb_image[0].save("lemons_ldm3d_rgb.jpg")
|
2825 |
+
depth_image[0].save("lemons_ldm3d_depth.png")
|
2826 |
|
2827 |
# Upscale the previous output to a resolution of (1024, 1024)
|
2828 |
|
|
|
2830 |
|
2831 |
pipe_ldm3d_upscale.to("cuda")
|
2832 |
|
2833 |
+
low_res_img = Image.open("lemons_ldm3d_rgb.jpg").convert("RGB")
|
2834 |
+
low_res_depth = Image.open("lemons_ldm3d_depth.png").convert("L")
|
2835 |
outputs = pipe_ldm3d_upscale(prompt="high quality high resolution uhd 4k image", rgb=low_res_img, depth=low_res_depth, num_inference_steps=50, target_res=[1024, 1024])
|
2836 |
|
2837 |
+
upscaled_rgb, upscaled_depth = outputs.rgb[0], outputs.depth[0]
|
2838 |
+
upscaled_rgb.save("upscaled_lemons_rgb.png")
|
2839 |
+
upscaled_depth.save("upscaled_lemons_depth.png")
|
2840 |
+
```
|
2841 |
|
2842 |
### ControlNet + T2I Adapter Pipeline
|
2843 |
|
2844 |
+
This pipeline combines both ControlNet and T2IAdapter into a single pipeline, where the forward pass is executed once.
|
2845 |
+
It receives `control_image` and `adapter_image`, as well as `controlnet_conditioning_scale` and `adapter_conditioning_scale`, for the ControlNet and Adapter modules, respectively. Whenever `adapter_conditioning_scale=0` or `controlnet_conditioning_scale=0`, it will act as a full ControlNet module or as a full T2IAdapter module, respectively.
|
2846 |
|
2847 |
```py
|
2848 |
import cv2
|
|
|
2905 |
adapter_conditioning_scale=strength,
|
2906 |
).images
|
2907 |
images[0].save("controlnet_and_adapter.png")
|
|
|
2908 |
```
|
2909 |
|
2910 |
### ControlNet + T2I Adapter + Inpainting Pipeline
|
|
|
2975 |
strength=0.7,
|
2976 |
).images
|
2977 |
images[0].save("controlnet_and_adapter_inpaint.png")
|
|
|
2978 |
```
|
2979 |
|
2980 |
### Regional Prompting Pipeline
|
2981 |
|
2982 |
+
This pipeline is a port of the [Regional Prompter extension](https://github.com/hako-mikan/sd-webui-regional-prompter) for [Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui) to `diffusers`.
|
2983 |
This code implements a pipeline for the Stable Diffusion model, enabling the division of the canvas into multiple regions, with different prompts applicable to each region. Users can specify regions in two ways: using `Cols` and `Rows` modes for grid-like divisions, or the `Prompt` mode for regions calculated based on prompts.
|
2984 |
|
2985 |

|
|
|
2990 |
|
2991 |
```py
|
2992 |
from examples.community.regional_prompting_stable_diffusion import RegionalPromptingStableDiffusionPipeline
|
2993 |
+
|
2994 |
pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae=vae)
|
2995 |
|
2996 |
rp_args = {
|
|
|
2998 |
"div": "1;1;1"
|
2999 |
}
|
3000 |
|
3001 |
+
prompt = """
|
3002 |
green hair twintail BREAK
|
3003 |
red blouse BREAK
|
3004 |
blue skirt
|
|
|
3008 |
prompt=prompt,
|
3009 |
negative_prompt=negative_prompt,
|
3010 |
guidance_scale=7.5,
|
3011 |
+
height=768,
|
3012 |
+
width=512,
|
3013 |
+
num_inference_steps=20,
|
3014 |
+
num_images_per_prompt=1,
|
3015 |
+
rp_args=rp_args
|
3016 |
+
).images
|
3017 |
|
3018 |
time = time.strftime(r"%Y%m%d%H%M%S")
|
3019 |
i = 1
|
|
|
3038 |
|
3039 |
### 2-Dimentional division
|
3040 |
|
3041 |
+
The prompt consists of instructions separated by the term `BREAK` and is assigned to different regions of a two-dimensional space. The image is initially split in the main splitting direction, which in this case is rows, due to the presence of a single semicolon `;`, dividing the space into an upper and a lower section. Additional sub-splitting is then applied, indicated by commas. The upper row is split into ratios of `2:1:1`, while the lower row is split into a ratio of `4:6`. Rows themselves are split in a `1:2` ratio. According to the reference image, the blue sky is designated as the first region, green hair as the second, the bookshelf as the third, and so on, in a sequence based on their position from the top left. The terrarium is placed on the desk in the fourth region, and the orange dress and sofa are in the fifth region, conforming to their respective splits.
|
3042 |
|
3043 |
+
```py
|
3044 |
rp_args = {
|
3045 |
"mode":"rows",
|
3046 |
"div": "1,2,1,1;2,4,6"
|
3047 |
}
|
3048 |
|
3049 |
+
prompt = """
|
3050 |
blue sky BREAK
|
3051 |
green hair BREAK
|
3052 |
book shelf BREAK
|
3053 |
+
terrarium on the desk BREAK
|
3054 |
orange dress and sofa
|
3055 |
"""
|
3056 |
```
|
|
|
3059 |
|
3060 |
### Prompt Mode
|
3061 |
|
3062 |
+
There are limitations to methods of specifying regions in advance. This is because specifying regions can be a hindrance when designating complex shapes or dynamic compositions. In the region specified by the prompt, the region is determined after the image generation has begun. This allows us to accommodate compositions and complex regions.
|
3063 |
For further infomagen, see [here](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/main/prompt_en.md).
|
3064 |
|
3065 |
+
### Syntax
|
3066 |
|
3067 |
```
|
3068 |
baseprompt target1 target2 BREAK
|
|
|
3084 |
|
3085 |
In this example, masks are calculated for shirt, tie, skirt, and color prompts are specified only for those regions.
|
3086 |
|
3087 |
+
```py
|
3088 |
rp_args = {
|
3089 |
+
"mode": "prompt-ex",
|
3090 |
+
"save_mask": True,
|
3091 |
"th": "0.4,0.6,0.6",
|
3092 |
}
|
3093 |
|
3094 |
+
prompt = """
|
3095 |
a girl in street with shirt, tie, skirt BREAK
|
3096 |
red, shirt BREAK
|
3097 |
green, tie BREAK
|
|
|
3101 |
|
3102 |

|
3103 |
|
3104 |
+
### Threshold
|
3105 |
|
3106 |
The threshold used to determine the mask created by the prompt. This can be set as many times as there are masks, as the range varies widely depending on the target prompt. If multiple regions are used, enter them separated by commas. For example, hair tends to be ambiguous and requires a small value, while face tends to be large and requires a small value. These should be ordered by BREAK.
|
3107 |
|
|
|
3120 |
|
3121 |
### Accuracy
|
3122 |
|
3123 |
+
In the case of a 512x512 image, Attention mode reduces the size of the region to about 8x8 pixels deep in the U-Net, so that small regions get mixed up; Latent mode calculates 64*64, so that the region is exact.
|
3124 |
|
3125 |
```
|
3126 |
girl hair twintail frills,ribbons, dress, face BREAK
|
|
|
3133 |
|
3134 |
### Use common prompt
|
3135 |
|
3136 |
+
You can attach the prompt up to ADDCOMM to all prompts by separating it first with ADDCOMM. This is useful when you want to include elements common to all regions. For example, when generating pictures of three people with different appearances, it's necessary to include the instruction of 'three people' in all regions. It's also useful when inserting quality tags and other things. "For example, if you write as follows:
|
3137 |
|
3138 |
```
|
3139 |
best quality, 3persons in garden, ADDCOMM
|
|
|
3156 |
|
3157 |
### Parameters
|
3158 |
|
3159 |
+
To activate Regional Prompter, it is necessary to enter settings in `rp_args`. The items that can be set are as follows. `rp_args` is a dictionary type.
|
3160 |
|
3161 |
### Input Parameters
|
3162 |
|
3163 |
Parameters are specified through the `rp_arg`(dictionary type).
|
3164 |
|
3165 |
+
```py
|
3166 |
rp_args = {
|
3167 |
"mode":"rows",
|
3168 |
"div": "1;1;1"
|
3169 |
}
|
3170 |
|
3171 |
+
pipe(prompt=prompt, rp_args=rp_args)
|
3172 |
```
|
3173 |
|
3174 |
### Required Parameters
|
3175 |
|
3176 |
+
- `mode`: Specifies the method for defining regions. Choose from `Cols`, `Rows`, `Prompt`, or `Prompt-Ex`. This parameter is case-insensitive.
|
3177 |
- `divide`: Used in `Cols` and `Rows` modes. Details on how to specify this are provided under the respective `Cols` and `Rows` sections.
|
3178 |
- `th`: Used in `Prompt` mode. The method of specification is detailed under the `Prompt` section.
|
3179 |
|
|
|
3187 |
|
3188 |
- Reference paper
|
3189 |
|
3190 |
+
```bibtex
|
3191 |
@article{chung2022diffusion,
|
3192 |
title={Diffusion posterior sampling for general noisy inverse problems},
|
3193 |
author={Chung, Hyungjin and Kim, Jeongsol and Mccann, Michael T and Klasky, Marc L and Ye, Jong Chul},
|
|
|
3199 |
- This pipeline allows zero-shot conditional sampling from the posterior distribution $p(x|y)$, given observation on $y$, unconditional generative model $p(x)$ and differentiable operator $y=f(x)$.
|
3200 |
|
3201 |
- For example, $f(.)$ can be downsample operator, then $y$ is a downsampled image, and the pipeline becomes a super-resolution pipeline.
|
3202 |
+
- To use this pipeline, you need to know your operator $f(.)$ and corrupted image $y$, and pass them during the call. For example, as in the main function of `dps_pipeline.py`, you need to first define the Gaussian blurring operator $f(.)$. The operator should be a callable `nn.Module`, with all the parameter gradient disabled:
|
3203 |
|
3204 |
```python
|
3205 |
import torch.nn.functional as F
|
|
|
3229 |
def weights_init(self):
|
3230 |
if self.blur_type == "gaussian":
|
3231 |
n = np.zeros((self.kernel_size, self.kernel_size))
|
3232 |
+
n[self.kernel_size // 2, self.kernel_size // 2] = 1
|
3233 |
k = scipy.ndimage.gaussian_filter(n, sigma=self.std)
|
3234 |
k = torch.from_numpy(k)
|
3235 |
self.k = k
|
|
|
3259 |
self.conv.update_weights(self.kernel.type(torch.float32))
|
3260 |
|
3261 |
for param in self.parameters():
|
3262 |
+
param.requires_grad = False
|
3263 |
|
3264 |
def forward(self, data, **kwargs):
|
3265 |
return self.conv(data)
|
|
|
3296 |
- 
|
3297 |
- Gaussian blurred image:
|
3298 |
- 
|
3299 |
+
- You can download those images to run the example on your own.
|
3300 |
|
3301 |
- Next, we need to define a loss function used for diffusion posterior sample. For most of the cases, the RMSE is fine:
|
3302 |
|
|
|
3305 |
return torch.sqrt(torch.sum((yhat-y)**2))
|
3306 |
```
|
3307 |
|
3308 |
+
- And next, as any other diffusion models, we need the score estimator and scheduler. As we are working with $256x256$ face images, we use ddpm-celebahq-256:
|
3309 |
|
3310 |
```python
|
3311 |
# set up scheduler
|
|
|
3322 |
# finally, the pipeline
|
3323 |
dpspipe = DPSPipeline(model, scheduler)
|
3324 |
image = dpspipe(
|
3325 |
+
measurement=measurement,
|
3326 |
+
operator=operator,
|
3327 |
+
loss_fn=RMSELoss,
|
3328 |
+
zeta=1.0,
|
3329 |
).images[0]
|
3330 |
image.save("dps_generated_image.png")
|
3331 |
```
|
3332 |
|
3333 |
+
- The `zeta` is a hyperparameter that is in range of $[0,1]$. It needs to be tuned for best effect. By setting `zeta=1`, you should be able to have the reconstructed result:
|
3334 |
- Reconstructed image:
|
3335 |
- 
|
3336 |
|
3337 |
- The reconstruction is perceptually similar to the source image, but different in details.
|
3338 |
+
- In `dps_pipeline.py`, we also provide a super-resolution example, which should produce:
|
3339 |
- Downsampled image:
|
3340 |
- 
|
3341 |
- Reconstructed image:
|
|
|
3347 |
|
3348 |
```py
|
3349 |
import torch
|
3350 |
+
from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter, DiffusionPipeline, DPMSolverMultistepScheduler
|
3351 |
+
from diffusers.utils import export_to_gif
|
|
|
3352 |
from PIL import Image
|
3353 |
|
3354 |
motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
|
|
|
3363 |
controlnet=controlnet,
|
3364 |
vae=vae,
|
3365 |
custom_pipeline="pipeline_animatediff_controlnet",
|
3366 |
+
torch_dtype=torch.float16,
|
3367 |
+
).to(device="cuda")
|
3368 |
pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
|
3369 |
model_id, subfolder="scheduler", beta_schedule="linear", clip_sample=False, timestep_spacing="linspace", steps_offset=1
|
3370 |
)
|
|
|
3385 |
num_inference_steps=20,
|
3386 |
).frames[0]
|
3387 |
|
|
|
3388 |
export_to_gif(result.frames[0], "result.gif")
|
3389 |
```
|
3390 |
|
|
|
3409 |
|
3410 |
```python
|
3411 |
import torch
|
3412 |
+
from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter, DiffusionPipeline, DPMSolverMultistepScheduler
|
3413 |
+
from diffusers.utils import export_to_gif
|
|
|
3414 |
from PIL import Image
|
3415 |
|
3416 |
motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
|
|
|
3426 |
controlnet=[controlnet1, controlnet2],
|
3427 |
vae=vae,
|
3428 |
custom_pipeline="pipeline_animatediff_controlnet",
|
3429 |
+
torch_dtype=torch.float16,
|
3430 |
+
).to(device="cuda")
|
3431 |
pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
|
3432 |
model_id, subfolder="scheduler", clip_sample=False, timestep_spacing="linspace", steps_offset=1, beta_schedule="linear",
|
3433 |
)
|
|
|
3474 |
num_inference_steps=20,
|
3475 |
)
|
3476 |
|
|
|
3477 |
export_to_gif(result.frames[0], "result.gif")
|
3478 |
```
|
3479 |
|
|
|
3602 |
output = pipe(prompt, image, mask_image, source_points, target_points)
|
3603 |
output_image = PIL.Image.fromarray(output)
|
3604 |
output_image.save("./output.png")
|
|
|
3605 |
```
|
3606 |
|
3607 |
### Instaflow Pipeline
|
|
|
3650 |
|
3651 |
- Reference paper
|
3652 |
|
3653 |
+
```bibtex
|
3654 |
+
@article{hertz2022prompt,
|
3655 |
+
title={Prompt-to-prompt image editing with cross attention control},
|
3656 |
+
author={Hertz, Amir and Mokady, Ron and Tenenbaum, Jay and Aberman, Kfir and Pritch, Yael and Cohen-Or, Daniel},
|
3657 |
+
booktitle={arXiv preprint arXiv:2208.01626},
|
3658 |
+
year={2022}
|
3659 |
```}
|
3660 |
|
3661 |
```py
|
3662 |
+
from diffusers import DDIMScheduler
|
3663 |
from examples.community.pipeline_null_text_inversion import NullTextPipeline
|
3664 |
import torch
|
3665 |
|
|
|
3666 |
device = "cuda"
|
3667 |
# Provide invert_prompt and the image for null-text optimization.
|
3668 |
invert_prompt = "A lying cat"
|
|
|
3674 |
# or different if editing.
|
3675 |
prompt = "A lying dog"
|
3676 |
|
3677 |
+
# Float32 is essential to a well optimization
|
3678 |
model_path = "runwayml/stable-diffusion-v1-5"
|
3679 |
scheduler = DDIMScheduler(num_train_timesteps=1000, beta_start=0.00085, beta_end=0.0120, beta_schedule="scaled_linear")
|
3680 |
+
pipeline = NullTextPipeline.from_pretrained(model_path, scheduler=scheduler, torch_dtype=torch.float32).to(device)
|
3681 |
|
3682 |
+
# Saves the inverted_latent to save time
|
3683 |
+
inverted_latent, uncond = pipeline.invert(input_image, invert_prompt, num_inner_steps=10, early_stop_epsilon=1e-5, num_inference_steps=steps)
|
3684 |
pipeline(prompt, uncond, inverted_latent, guidance_scale=7.5, num_inference_steps=steps).images[0].save(input_image+".output.jpg")
|
3685 |
```
|
3686 |
|
|
|
3737 |
controlnet = ControlNetModel.from_pretrained(
|
3738 |
"lllyasviel/sd-controlnet-canny").to('cuda')
|
3739 |
|
3740 |
+
# You can use any finetuned SD here
|
3741 |
pipe = DiffusionPipeline.from_pretrained(
|
3742 |
"runwayml/stable-diffusion-v1-5", controlnet=controlnet, custom_pipeline='rerender_a_video').to('cuda')
|
3743 |
|
|
|
3779 |
from typing import List
|
3780 |
|
3781 |
import torch
|
3782 |
+
from diffusers import DiffusionPipeline
|
3783 |
from PIL import Image
|
3784 |
|
3785 |
model_id = "a-r-r-o-w/dreamshaper-xl-turbo"
|
|
|
3848 |
image=image,
|
3849 |
prompt="A snail moving on the ground",
|
3850 |
strength=0.8,
|
3851 |
+
latent_interpolation_method="slerp", # can be lerp, slerp, or your own callback
|
3852 |
)
|
3853 |
frames = output.frames[0]
|
3854 |
export_to_gif(frames, "animation.gif")
|
|
|
3858 |
|
3859 |
IP Adapter FaceID is an experimental IP Adapter model that uses image embeddings generated by `insightface`, so no image encoder needs to be loaded.
|
3860 |
You need to install `insightface` and all its requirements to use this model.
|
3861 |
+
You must pass the image embedding tensor as `image_embeds` to the `DiffusionPipeline` instead of `ip_adapter_image`.
|
3862 |
You can find more results [here](https://github.com/huggingface/diffusers/pull/6276).
|
3863 |
|
3864 |
```py
|
|
|
3865 |
import torch
|
3866 |
from diffusers.utils import load_image
|
3867 |
import cv2
|
|
|
3891 |
pipeline.to("cuda")
|
3892 |
|
3893 |
generator = torch.Generator(device="cpu").manual_seed(42)
|
3894 |
+
num_images = 2
|
3895 |
|
3896 |
image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png")
|
3897 |
|
|
|
3914 |
|
3915 |
### InstantID Pipeline
|
3916 |
|
3917 |
+
InstantID is a new state-of-the-art tuning-free method to achieve ID-Preserving generation with only single image, supporting various downstream tasks. For any usage question, please refer to the [official implementation](https://github.com/InstantID/InstantID).
|
3918 |
|
3919 |
```py
|
3920 |
+
# !pip install diffusers opencv-python transformers accelerate insightface
|
3921 |
import diffusers
|
3922 |
from diffusers.utils import load_image
|
3923 |
+
from diffusers import ControlNetModel
|
3924 |
|
3925 |
import cv2
|
3926 |
import torch
|
|
|
3938 |
# prepare models under ./checkpoints
|
3939 |
# https://huggingface.co/InstantX/InstantID
|
3940 |
from huggingface_hub import hf_hub_download
|
3941 |
+
|
3942 |
hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/config.json", local_dir="./checkpoints")
|
3943 |
hf_hub_download(repo_id="InstantX/InstantID", filename="ControlNetModel/diffusion_pytorch_model.safetensors", local_dir="./checkpoints")
|
3944 |
hf_hub_download(repo_id="InstantX/InstantID", filename="ip-adapter.bin", local_dir="./checkpoints")
|
3945 |
|
3946 |
+
face_adapter = './checkpoints/ip-adapter.bin'
|
3947 |
+
controlnet_path = './checkpoints/ControlNetModel'
|
3948 |
|
3949 |
# load IdentityNet
|
3950 |
controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16)
|
|
|
3955 |
controlnet=controlnet,
|
3956 |
torch_dtype=torch.float16
|
3957 |
)
|
3958 |
+
pipe.to("cuda")
|
3959 |
|
3960 |
# load adapter
|
3961 |
pipe.load_ip_adapter_instantid(face_adapter)
|
|
|
4022 |
import torch
|
4023 |
import numpy as np
|
4024 |
|
4025 |
+
from diffusers import ControlNetModel, DDIMScheduler, DiffusionPipeline
|
4026 |
import sys
|
4027 |
+
|
4028 |
gmflow_dir = "/path/to/gmflow"
|
4029 |
sys.path.insert(0, gmflow_dir)
|
4030 |
|
|
|
4052 |
input_video_path = 'https://github.com/williamyang1991/FRESCO/raw/main/data/car-turn.mp4'
|
4053 |
output_video_path = 'car.gif'
|
4054 |
|
4055 |
+
# You can use any finetuned SD here
|
4056 |
model_path = 'SG161222/Realistic_Vision_V2.0'
|
4057 |
|
4058 |
prompt = 'a red car turns in the winter'
|
|
|
4097 |
|
4098 |
output_frames[0].save(output_video_path, save_all=True,
|
4099 |
append_images=output_frames[1:], duration=100, loop=0)
|
|
|
4100 |
```
|
4101 |
|
4102 |
# Perturbed-Attention Guidance
|
4103 |
|
4104 |
[Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)
|
4105 |
|
4106 |
+
This implementation is based on [Diffusers](https://huggingface.co/docs/diffusers/index). `StableDiffusionPAGPipeline` is a modification of `StableDiffusionPipeline` to support Perturbed-Attention Guidance (PAG).
|
4107 |
|
4108 |
## Example Usage
|
4109 |
|
|
|
4123 |
torch_dtype=torch.float16
|
4124 |
)
|
4125 |
|
4126 |
+
device = "cuda"
|
4127 |
pipe = pipe.to(device)
|
4128 |
|
4129 |
pag_scale = 5.0
|
4130 |
pag_applied_layers_index = ['m0']
|
4131 |
|
4132 |
batch_size = 4
|
4133 |
+
seed = 10
|
4134 |
|
4135 |
base_dir = "./results/"
|
4136 |
grid_dir = base_dir + "/pag" + str(pag_scale) + "/"
|
|
|
4140 |
|
4141 |
set_seed(seed)
|
4142 |
|
4143 |
+
latent_input = randn_tensor(shape=(batch_size,4,64,64), generator=None, device=device, dtype=torch.float16)
|
4144 |
|
4145 |
output_baseline = pipe(
|
4146 |
"",
|
|
|
4172 |
|
4173 |
## PAG Parameters
|
4174 |
|
4175 |
+
`pag_scale` : guidance scale of PAG (ex: 5.0)
|
4176 |
|
4177 |
+
`pag_applied_layers_index` : index of the layer to apply perturbation (ex: ['m0'])
|
main/lpw_stable_diffusion_xl.py
CHANGED
@@ -2,7 +2,7 @@
|
|
2 |
# A SDXL pipeline can take unlimited weighted prompt
|
3 |
#
|
4 |
# Author: Andrew Zhu
|
5 |
-
#
|
6 |
# Medium: https://medium.com/@xhinker
|
7 |
## -----------------------------------------------------------
|
8 |
|
|
|
2 |
# A SDXL pipeline can take unlimited weighted prompt
|
3 |
#
|
4 |
# Author: Andrew Zhu
|
5 |
+
# GitHub: https://github.com/xhinker
|
6 |
# Medium: https://medium.com/@xhinker
|
7 |
## -----------------------------------------------------------
|
8 |
|