NadaGh's picture
End of training
dde5d93 verified
|
raw
history blame
12.5 kB
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# DiffEdit
[[open-in-colab]]
์ด๋ฏธ์ง€ ํŽธ์ง‘์„ ํ•˜๋ ค๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ํŽธ์ง‘ํ•  ์˜์—ญ์˜ ๋งˆ์Šคํฌ๋ฅผ ์ œ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. DiffEdit๋Š” ํ…์ŠคํŠธ ์ฟผ๋ฆฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋งˆ์Šคํฌ๋ฅผ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋ฏ€๋กœ ์ด๋ฏธ์ง€ ํŽธ์ง‘ ์†Œํ”„ํŠธ์›จ์–ด ์—†์ด๋„ ๋งˆ์Šคํฌ๋ฅผ ๋งŒ๋“ค๊ธฐ๊ฐ€ ์ „๋ฐ˜์ ์œผ๋กœ ๋” ์‰ฌ์›Œ์ง‘๋‹ˆ๋‹ค. DiffEdit ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์„ธ ๋‹จ๊ณ„๋กœ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค:
1. Diffusion ๋ชจ๋ธ์ด ์ผ๋ถ€ ์ฟผ๋ฆฌ ํ…์ŠคํŠธ์™€ ์ฐธ์กฐ ํ…์ŠคํŠธ๋ฅผ ์กฐ๊ฑด๋ถ€๋กœ ์ด๋ฏธ์ง€์˜ ๋…ธ์ด์ฆˆ๋ฅผ ์ œ๊ฑฐํ•˜์—ฌ ์ด๋ฏธ์ง€์˜ ์—ฌ๋Ÿฌ ์˜์—ญ์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋…ธ์ด์ฆˆ ์ถ”์ •์น˜๋ฅผ ์ƒ์„ฑํ•˜๊ณ , ๊ทธ ์ฐจ์ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ฟผ๋ฆฌ ํ…์ŠคํŠธ์™€ ์ผ์น˜ํ•˜๋„๋ก ์ด๋ฏธ์ง€์˜ ์–ด๋Š ์˜์—ญ์„ ๋ณ€๊ฒฝํ•ด์•ผ ํ•˜๋Š”์ง€ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ๋งˆ์Šคํฌ๋ฅผ ์ถ”๋ก ํ•ฉ๋‹ˆ๋‹ค.
2. ์ž…๋ ฅ ์ด๋ฏธ์ง€๊ฐ€ DDIM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž ์žฌ ๊ณต๊ฐ„์œผ๋กœ ์ธ์ฝ”๋”ฉ๋ฉ๋‹ˆ๋‹ค.
3. ๋งˆ์Šคํฌ ์™ธ๋ถ€์˜ ํ”ฝ์…€์ด ์ž…๋ ฅ ์ด๋ฏธ์ง€์™€ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€๋˜๋„๋ก ๋งˆ์Šคํฌ๋ฅผ ๊ฐ€์ด๋“œ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ ์ฟผ๋ฆฌ์— ์กฐ๊ฑด์ด ์ง€์ •๋œ diffusion ๋ชจ๋ธ๋กœ latents๋ฅผ ๋””์ฝ”๋”ฉํ•ฉ๋‹ˆ๋‹ค.
์ด ๊ฐ€์ด๋“œ์—์„œ๋Š” ๋งˆ์Šคํฌ๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋งŒ๋“ค์ง€ ์•Š๊ณ  DiffEdit๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ํŽธ์ง‘ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
์‹œ์ž‘ํ•˜๊ธฐ ์ „์— ๋‹ค์Œ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๊ฐ€ ์„ค์น˜๋˜์–ด ์žˆ๋Š”์ง€ ํ™•์ธํ•˜์„ธ์š”:
```py
# Colab์—์„œ ํ•„์š”ํ•œ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์„ค์น˜ํ•˜๊ธฐ ์œ„ํ•ด ์ฃผ์„์„ ์ œ์™ธํ•˜์„ธ์š”
#!pip install -q diffusers transformers accelerate
```
[`StableDiffusionDiffEditPipeline`]์—๋Š” ์ด๋ฏธ์ง€ ๋งˆ์Šคํฌ์™€ ๋ถ€๋ถ„์ ์œผ๋กœ ๋ฐ˜์ „๋œ latents ์ง‘ํ•ฉ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ๋งˆ์Šคํฌ๋Š” [`~StableDiffusionDiffEditPipeline.generate_mask`] ํ•จ์ˆ˜์—์„œ ์ƒ์„ฑ๋˜๋ฉฐ, ๋‘ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ์ธ `source_prompt`์™€ `target_prompt`๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด ๋งค๊ฐœ๋ณ€์ˆ˜๋Š” ์ด๋ฏธ์ง€์—์„œ ๋ฌด์—‡์„ ํŽธ์ง‘ํ• ์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, *๊ณผ์ผ* ํ•œ ๊ทธ๋ฆ‡์„ *๋ฐฐ* ํ•œ ๊ทธ๋ฆ‡์œผ๋กœ ๋ณ€๊ฒฝํ•˜๋ ค๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ•˜์„ธ์š”:
```py
source_prompt = "a bowl of fruits"
target_prompt = "a bowl of pears"
```
๋ถ€๋ถ„์ ์œผ๋กœ ๋ฐ˜์ „๋œ latents๋Š” [`~StableDiffusionDiffEditPipeline.invert`] ํ•จ์ˆ˜์—์„œ ์ƒ์„ฑ๋˜๋ฉฐ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ด๋ฏธ์ง€๋ฅผ ์„ค๋ช…ํ•˜๋Š” `prompt` ๋˜๋Š” *์บก์…˜*์„ ํฌํ•จํ•˜๋Š” ๊ฒƒ์ด inverse latent sampling ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€์ด๋“œํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์บก์…˜์€ ์ข…์ข… `source_prompt`๊ฐ€ ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋‹ค๋ฅธ ํ…์ŠคํŠธ ์„ค๋ช…์œผ๋กœ ์ž์œ ๋กญ๊ฒŒ ์‹คํ—˜ํ•ด ๋ณด์„ธ์š”!
ํŒŒ์ดํ”„๋ผ์ธ, ์Šค์ผ€์ค„๋Ÿฌ, ์—ญ ์Šค์ผ€์ค„๋Ÿฌ๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๊ณ  ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋ช‡ ๊ฐ€์ง€ ์ตœ์ ํ™”๋ฅผ ํ™œ์„ฑํ™”ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค:
```py
import torch
from diffusers import DDIMScheduler, DDIMInverseScheduler, StableDiffusionDiffEditPipeline
pipeline = StableDiffusionDiffEditPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1",
torch_dtype=torch.float16,
safety_checker=None,
use_safetensors=True,
)
pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
pipeline.enable_model_cpu_offload()
pipeline.enable_vae_slicing()
```
์ˆ˜์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ด๋ฏธ์ง€๋ฅผ ๋ถˆ๋Ÿฌ์˜ต๋‹ˆ๋‹ค:
```py
from diffusers.utils import load_image, make_image_grid
img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
raw_image = load_image(img_url).resize((768, 768))
raw_image
```
์ด๋ฏธ์ง€ ๋งˆ์Šคํฌ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด [`~StableDiffusionDiffEditPipeline.generate_mask`] ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€์—์„œ ํŽธ์ง‘ํ•  ๋‚ด์šฉ์„ ์ง€์ •ํ•˜๊ธฐ ์œ„ํ•ด `source_prompt`์™€ `target_prompt`๋ฅผ ์ „๋‹ฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค:
```py
from PIL import Image
source_prompt = "a bowl of fruits"
target_prompt = "a basket of pears"
mask_image = pipeline.generate_mask(
image=raw_image,
source_prompt=source_prompt,
target_prompt=target_prompt,
)
Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L").resize((768, 768))
```
๋‹ค์Œ์œผ๋กœ, ๋ฐ˜์ „๋œ latents๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์ด๋ฏธ์ง€๋ฅผ ๋ฌ˜์‚ฌํ•˜๋Š” ์บก์…˜์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค:
```py
inv_latents = pipeline.invert(prompt=source_prompt, image=raw_image).latents
```
๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด๋ฏธ์ง€ ๋งˆ์Šคํฌ์™€ ๋ฐ˜์ „๋œ latents๋ฅผ ํŒŒ์ดํ”„๋ผ์ธ์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. `target_prompt`๋Š” ์ด์ œ `prompt`๊ฐ€ ๋˜๋ฉฐ, `source_prompt`๋Š” `negative_prompt`๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
```py
output_image = pipeline(
prompt=target_prompt,
mask_image=mask_image,
image_latents=inv_latents,
negative_prompt=source_prompt,
).images[0]
mask_image = Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L").resize((768, 768))
make_image_grid([raw_image, mask_image, output_image], rows=1, cols=3)
```
<div class="flex gap-4">
<div>
<img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">original image</figcaption>
</div>
<div>
<img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/blob/main/assets/target.png?raw=true"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">edited image</figcaption>
</div>
</div>
## Source์™€ target ์ž„๋ฒ ๋”ฉ ์ƒ์„ฑํ•˜๊ธฐ
Source์™€ target ์ž„๋ฒ ๋”ฉ์€ ์ˆ˜๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋Œ€์‹  [Flan-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Flan-T5 ๋ชจ๋ธ๊ณผ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ๐Ÿค— Transformers ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ ๋ถˆ๋Ÿฌ์˜ต๋‹ˆ๋‹ค:
```py
import torch
from transformers import AutoTokenizer, T5ForConditionalGeneration
tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large")
model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large", device_map="auto", torch_dtype=torch.float16)
```
๋ชจ๋ธ์— ํ”„๋กฌํ”„ํŠธํ•  source์™€ target ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ดˆ๊ธฐ ํ…์ŠคํŠธ๋“ค์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
```py
source_concept = "bowl"
target_concept = "basket"
source_text = f"Provide a caption for images containing a {source_concept}. "
"The captions should be in English and should be no longer than 150 characters."
target_text = f"Provide a caption for images containing a {target_concept}. "
"The captions should be in English and should be no longer than 150 characters."
```
๋‹ค์Œ์œผ๋กœ, ํ”„๋กฌํ”„ํŠธ๋“ค์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์œ ํ‹ธ๋ฆฌํ‹ฐ ํ•จ์ˆ˜๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
```py
@torch.no_grad()
def generate_prompts(input_prompt):
input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids.to("cuda")
outputs = model.generate(
input_ids, temperature=0.8, num_return_sequences=16, do_sample=True, max_new_tokens=128, top_k=10
)
return tokenizer.batch_decode(outputs, skip_special_tokens=True)
source_prompts = generate_prompts(source_text)
target_prompts = generate_prompts(target_text)
print(source_prompts)
print(target_prompts)
```
<Tip>
๋‹ค์–‘ํ•œ ํ’ˆ์งˆ์˜ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ „๋žต์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด๋ ค๋ฉด [์ƒ์„ฑ ์ „๋žต](https://huggingface.co/docs/transformers/main/en/generation_strategies) ๊ฐ€์ด๋“œ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
</Tip>
ํ…์ŠคํŠธ ์ธ์ฝ”๋”ฉ์„ ์œ„ํ•ด [`StableDiffusionDiffEditPipeline`]์—์„œ ์‚ฌ์šฉํ•˜๋Š” ํ…์ŠคํŠธ ์ธ์ฝ”๋” ๋ชจ๋ธ์„ ๋ถˆ๋Ÿฌ์˜ต๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ์ธ์ฝ”๋”๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค:
```py
import torch
from diffusers import StableDiffusionDiffEditPipeline
pipeline = StableDiffusionDiffEditPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16, use_safetensors=True
)
pipeline.enable_model_cpu_offload()
pipeline.enable_vae_slicing()
@torch.no_grad()
def embed_prompts(sentences, tokenizer, text_encoder, device="cuda"):
embeddings = []
for sent in sentences:
text_inputs = tokenizer(
sent,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
prompt_embeds = text_encoder(text_input_ids.to(device), attention_mask=None)[0]
embeddings.append(prompt_embeds)
return torch.concatenate(embeddings, dim=0).mean(dim=0).unsqueeze(0)
source_embeds = embed_prompts(source_prompts, pipeline.tokenizer, pipeline.text_encoder)
target_embeds = embed_prompts(target_prompts, pipeline.tokenizer, pipeline.text_encoder)
```
๋งˆ์ง€๋ง‰์œผ๋กœ, ์ž„๋ฒ ๋”ฉ์„ [`~StableDiffusionDiffEditPipeline.generate_mask`] ๋ฐ [`~StableDiffusionDiffEditPipeline.invert`] ํ•จ์ˆ˜์™€ ํŒŒ์ดํ”„๋ผ์ธ์— ์ „๋‹ฌํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค:
```diff
from diffusers import DDIMInverseScheduler, DDIMScheduler
from diffusers.utils import load_image, make_image_grid
from PIL import Image
pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
raw_image = load_image(img_url).resize((768, 768))
mask_image = pipeline.generate_mask(
image=raw_image,
- source_prompt=source_prompt,
- target_prompt=target_prompt,
+ source_prompt_embeds=source_embeds,
+ target_prompt_embeds=target_embeds,
)
inv_latents = pipeline.invert(
- prompt=source_prompt,
+ prompt_embeds=source_embeds,
image=raw_image,
).latents
output_image = pipeline(
mask_image=mask_image,
image_latents=inv_latents,
- prompt=target_prompt,
- negative_prompt=source_prompt,
+ prompt_embeds=target_embeds,
+ negative_prompt_embeds=source_embeds,
).images[0]
mask_image = Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L")
make_image_grid([raw_image, mask_image, output_image], rows=1, cols=3)
```
## ๋ฐ˜์ „์„ ์œ„ํ•œ ์บก์…˜ ์ƒ์„ฑํ•˜๊ธฐ
`source_prompt`๋ฅผ ์บก์…˜์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋ถ€๋ถ„์ ์œผ๋กœ ๋ฐ˜์ „๋œ latents๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, [BLIP](https://huggingface.co/docs/transformers/model_doc/blip) ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์บก์…˜์„ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿค— Transformers ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ BLIP ๋ชจ๋ธ๊ณผ ํ”„๋กœ์„ธ์„œ๋ฅผ ๋ถˆ๋Ÿฌ์˜ต๋‹ˆ๋‹ค:
```py
import torch
from transformers import BlipForConditionalGeneration, BlipProcessor
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base", torch_dtype=torch.float16, low_cpu_mem_usage=True)
```
์ž…๋ ฅ ์ด๋ฏธ์ง€์—์„œ ์บก์…˜์„ ์ƒ์„ฑํ•˜๋Š” ์œ ํ‹ธ๋ฆฌํ‹ฐ ํ•จ์ˆ˜๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค:
```py
@torch.no_grad()
def generate_caption(images, caption_generator, caption_processor):
text = "a photograph of"
inputs = caption_processor(images, text, return_tensors="pt").to(device="cuda", dtype=caption_generator.dtype)
caption_generator.to("cuda")
outputs = caption_generator.generate(**inputs, max_new_tokens=128)
# ์บก์…˜ generator ์˜คํ”„๋กœ๋“œ
caption_generator.to("cpu")
caption = caption_processor.batch_decode(outputs, skip_special_tokens=True)[0]
return caption
```
์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ ๋ถˆ๋Ÿฌ์˜ค๊ณ  `generate_caption` ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•ด๋‹น ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์บก์…˜์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค:
```py
from diffusers.utils import load_image
img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
raw_image = load_image(img_url).resize((768, 768))
caption = generate_caption(raw_image, model, processor)
```
<div class="flex justify-center">
<figure>
<img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"/>
<figcaption class="text-center">generated caption: "a photograph of a bowl of fruit on a table"</figcaption>
</figure>
</div>
์ด์ œ ์บก์…˜์„ [`~StableDiffusionDiffEditPipeline.invert`] ํ•จ์ˆ˜์— ๋†“์•„ ๋ถ€๋ถ„์ ์œผ๋กœ ๋ฐ˜์ „๋œ latents๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค!