Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import torch
|
| 3 |
+
import gradio as gr
|
| 4 |
+
from tqdm import tqdm
|
| 5 |
+
from PIL import Image
|
| 6 |
+
import torch.nn.functional as F
|
| 7 |
+
from torchvision import transforms as tfms
|
| 8 |
+
from transformers import CLIPTextModel, CLIPTokenizer, logging
|
| 9 |
+
from diffusers import AutoencoderKL, LMSDiscreteScheduler, UNet2DConditionModel, DiffusionPipeline
|
| 10 |
+
|
| 11 |
+
torch_device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
| 12 |
+
if "mps" == torch_device: os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = "1"
|
| 13 |
+
|
| 14 |
+
# Load the pipeline
|
| 15 |
+
model_path = "CompVis/stable-diffusion-v1-4"
|
| 16 |
+
sd_pipeline = DiffusionPipeline.from_pretrained(
|
| 17 |
+
model_path,
|
| 18 |
+
low_cpu_mem_usage=True,
|
| 19 |
+
torch_dtype=torch.float32
|
| 20 |
+
).to(torch_device)
|
| 21 |
+
|
| 22 |
+
# Load textual inversions
|
| 23 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/illustration-style")
|
| 24 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/line-art")
|
| 25 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/hitokomoru-style-nao")
|
| 26 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/style-of-marc-allante")
|
| 27 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/midjourney-style")
|
| 28 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/hanfu-anime-style")
|
| 29 |
+
sd_pipeline.load_textual_inversion("sd-concepts-library/birb-style")
|
| 30 |
+
|
| 31 |
+
# Update style token dictionary
|
| 32 |
+
style_token_dict = {
|
| 33 |
+
"Illustration Style": '<illustration-style>',
|
| 34 |
+
"Line Art":'<line-art>',
|
| 35 |
+
"Hitokomoru Style":'<hitokomoru-style-nao>',
|
| 36 |
+
"Marc Allante": '<Marc_Allante>',
|
| 37 |
+
"Midjourney":'<midjourney-style>',
|
| 38 |
+
"Hanfu Anime": '<hanfu-anime-style>',
|
| 39 |
+
"Birb Style": '<birb-style>'
|
| 40 |
+
}
|
| 41 |
+
|
| 42 |
+
def apply_guidance(latents, guidance_method, loss_scale):
|
| 43 |
+
if guidance_method == 'Grayscale':
|
| 44 |
+
rgb = latents_to_pil(latents)[0]
|
| 45 |
+
gray = rgb.convert('L')
|
| 46 |
+
gray_latents = pil_to_latent(gray.convert('RGB'))
|
| 47 |
+
return latents + (gray_latents - latents) * loss_scale
|
| 48 |
+
elif guidance_method == 'Bright':
|
| 49 |
+
bright_latents = F.relu(latents) # Simple brightness increase
|
| 50 |
+
return latents + (bright_latents - latents) * loss_scale
|
| 51 |
+
elif guidance_method == 'Contrast':
|
| 52 |
+
mean = latents.mean()
|
| 53 |
+
contrast_latents = (latents - mean) * 2 + mean
|
| 54 |
+
return latents + (contrast_latents - latents) * loss_scale
|
| 55 |
+
elif guidance_method == 'Symmetry':
|
| 56 |
+
flipped_latents = torch.flip(latents, [3]) # Flip horizontally
|
| 57 |
+
return latents + (flipped_latents - latents) * loss_scale
|
| 58 |
+
elif guidance_method == 'Saturation':
|
| 59 |
+
rgb = latents_to_pil(latents)[0]
|
| 60 |
+
saturated = tfms.functional.adjust_saturation(tfms.ToTensor()(rgb), 2)
|
| 61 |
+
saturated_latents = pil_to_latent(tfms.ToPILImage()(saturated))
|
| 62 |
+
return latents + (saturated_latents - latents) * loss_scale
|
| 63 |
+
else:
|
| 64 |
+
return latents
|
| 65 |
+
|
| 66 |
+
def generate_with_guidance(prompt, num_inference_steps, guidance_scale, seed, guidance_method, loss_scale):
|
| 67 |
+
generator = torch.Generator(device=torch_device).manual_seed(seed)
|
| 68 |
+
|
| 69 |
+
# Get the text embeddings
|
| 70 |
+
text_input = sd_pipeline.tokenizer(prompt, padding="max_length", max_length=sd_pipeline.tokenizer.model_max_length, truncation=True, return_tensors="pt")
|
| 71 |
+
with torch.no_grad():
|
| 72 |
+
text_embeddings = sd_pipeline.text_encoder(text_input.input_ids.to(torch_device))[0]
|
| 73 |
+
|
| 74 |
+
# Set the timesteps
|
| 75 |
+
sd_pipeline.scheduler.set_timesteps(num_inference_steps)
|
| 76 |
+
|
| 77 |
+
# Prepare latents
|
| 78 |
+
latents = torch.randn(
|
| 79 |
+
(1, sd_pipeline.unet.in_channels, 64, 64),
|
| 80 |
+
generator=generator,
|
| 81 |
+
device=torch_device
|
| 82 |
+
)
|
| 83 |
+
latents = latents * sd_pipeline.scheduler.init_noise_sigma
|
| 84 |
+
|
| 85 |
+
# Denoising loop
|
| 86 |
+
for t in tqdm(sd_pipeline.scheduler.timesteps):
|
| 87 |
+
# Expand the latents for classifier-free guidance
|
| 88 |
+
latent_model_input = torch.cat([latents] * 2)
|
| 89 |
+
latent_model_input = sd_pipeline.scheduler.scale_model_input(latent_model_input, timestep=t)
|
| 90 |
+
|
| 91 |
+
# Predict the noise residual
|
| 92 |
+
with torch.no_grad():
|
| 93 |
+
noise_pred = sd_pipeline.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
| 94 |
+
|
| 95 |
+
# Perform guidance
|
| 96 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 97 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
| 98 |
+
|
| 99 |
+
# Apply custom guidance
|
| 100 |
+
latents = apply_guidance(latents, guidance_method, loss_scale / 10000) # Normalize loss_scale
|
| 101 |
+
|
| 102 |
+
# Compute the previous noisy sample x_t -> x_t-1
|
| 103 |
+
latents = sd_pipeline.scheduler.step(noise_pred, t, latents).prev_sample
|
| 104 |
+
|
| 105 |
+
# Scale and decode the image latents with vae
|
| 106 |
+
latents = 1 / 0.18215 * latents
|
| 107 |
+
with torch.no_grad():
|
| 108 |
+
image = sd_pipeline.vae.decode(latents).sample
|
| 109 |
+
|
| 110 |
+
# Convert to PIL Image
|
| 111 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
| 112 |
+
image = image.detach().cpu().permute(0, 2, 3, 1).numpy()
|
| 113 |
+
image = (image * 255).round().astype("uint8")[0]
|
| 114 |
+
image = Image.fromarray(image)
|
| 115 |
+
|
| 116 |
+
return image
|
| 117 |
+
|
| 118 |
+
def inference(text, style, inference_step, guidance_scale, seed, guidance_method, loss_scale):
|
| 119 |
+
prompt = text + " " + style_token_dict[style]
|
| 120 |
+
|
| 121 |
+
# Generate image with pipeline
|
| 122 |
+
image_pipeline = sd_pipeline(
|
| 123 |
+
prompt,
|
| 124 |
+
num_inference_steps=inference_step,
|
| 125 |
+
guidance_scale=guidance_scale,
|
| 126 |
+
generator=torch.Generator(device=torch_device).manual_seed(seed)
|
| 127 |
+
).images[0]
|
| 128 |
+
|
| 129 |
+
# Generate image with guidance
|
| 130 |
+
image_guide = generate_with_guidance(prompt, inference_step, guidance_scale, seed, guidance_method, loss_scale)
|
| 131 |
+
|
| 132 |
+
return image_pipeline, image_guide
|
| 133 |
+
|
| 134 |
+
title = "Generative with Textual Inversion and Guidance"
|
| 135 |
+
description = "A Gradio interface to infer Stable Diffusion and generate images with different art styles and guidance methods"
|
| 136 |
+
examples = [
|
| 137 |
+
["A majestic castle on a floating island", 'Illustration Style', 20, 7.5, 42, 'Grayscale', 200],
|
| 138 |
+
["A cyberpunk cityscape at night", 'Midjourney', 25, 8.0, 123, 'Contrast', 300]
|
| 139 |
+
]
|
| 140 |
+
|
| 141 |
+
demo = gr.Interface(inference,
|
| 142 |
+
inputs = [gr.Textbox(label="Prompt", type="text"),
|
| 143 |
+
gr.Dropdown(label="Style", choices=list(style_token_dict.keys()), value="Illustration Style"),
|
| 144 |
+
gr.Slider(1, 50, 10, step = 1, label="Inference steps"),
|
| 145 |
+
gr.Slider(1, 10, 7.5, step = 0.1, label="Guidance scale"),
|
| 146 |
+
gr.Slider(0, 10000, 42, step = 1, label="Seed"),
|
| 147 |
+
gr.Dropdown(label="Guidance method", choices=['Grayscale', 'Bright', 'Contrast',
|
| 148 |
+
'Symmetry', 'Saturation'], value="Grayscale"),
|
| 149 |
+
gr.Slider(100, 10000, 200, step = 100, label="Loss scale")],
|
| 150 |
+
outputs= [gr.Image(width=512, height=512, label="Generated art"),
|
| 151 |
+
gr.Image(width=512, height=512, label="Generated art with guidance")],
|
| 152 |
+
title=title,
|
| 153 |
+
description=description,
|
| 154 |
+
examples=examples)
|
| 155 |
+
|
| 156 |
+
demo.launch()
|