Spaces:
Running
on
Zero
Running
on
Zero
NightRaven109
commited on
Commit
•
7c89d3a
1
Parent(s):
6ecc7d4
Upload 2 files
Browse files- app.py +212 -0
- requirements.txt +15 -14
app.py
ADDED
@@ -0,0 +1,212 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import gradio as gr
|
4 |
+
import spaces
|
5 |
+
import numpy as np
|
6 |
+
from PIL import Image
|
7 |
+
import safetensors.torch
|
8 |
+
from huggingface_hub import hf_hub_download
|
9 |
+
from accelerate import Accelerator
|
10 |
+
from accelerate.utils import set_seed
|
11 |
+
from diffusers import (
|
12 |
+
AutoencoderKL,
|
13 |
+
DDPMScheduler,
|
14 |
+
UNet2DConditionModel,
|
15 |
+
)
|
16 |
+
from transformers import CLIPTextModel, CLIPTokenizer, CLIPImageProcessor
|
17 |
+
from models.controlnet import ControlNetModel
|
18 |
+
from pipelines.pipeline_ccsr import StableDiffusionControlNetPipeline
|
19 |
+
from myutils.wavelet_color_fix import wavelet_color_fix, adain_color_fix
|
20 |
+
|
21 |
+
# Initialize global variables for models
|
22 |
+
pipeline = None
|
23 |
+
generator = None
|
24 |
+
accelerator = None
|
25 |
+
|
26 |
+
@spaces.GPU
|
27 |
+
def initialize_models():
|
28 |
+
global pipeline, generator, accelerator
|
29 |
+
|
30 |
+
# Initialize accelerator
|
31 |
+
accelerator = Accelerator(
|
32 |
+
mixed_precision="fp16",
|
33 |
+
gradient_accumulation_steps=1
|
34 |
+
)
|
35 |
+
|
36 |
+
try:
|
37 |
+
# Download and load models with authentication token
|
38 |
+
scheduler = DDPMScheduler.from_pretrained(
|
39 |
+
"NightRaven109/CCSRModels",
|
40 |
+
subfolder="stable-diffusion-2-1-base/scheduler",
|
41 |
+
use_auth_token=os.environ['Read']
|
42 |
+
)
|
43 |
+
|
44 |
+
text_encoder = CLIPTextModel.from_pretrained(
|
45 |
+
"NightRaven109/CCSRModels",
|
46 |
+
subfolder="stable-diffusion-2-1-base/text_encoder",
|
47 |
+
use_auth_token=os.environ['Read']
|
48 |
+
)
|
49 |
+
|
50 |
+
tokenizer = CLIPTokenizer.from_pretrained(
|
51 |
+
"NightRaven109/CCSRModels",
|
52 |
+
subfolder="stable-diffusion-2-1-base/tokenizer",
|
53 |
+
use_auth_token=os.environ['Read']
|
54 |
+
)
|
55 |
+
|
56 |
+
feature_extractor = CLIPImageProcessor.from_pretrained(
|
57 |
+
"NightRaven109/CCSRModels",
|
58 |
+
subfolder="stable-diffusion-2-1-base/feature_extractor",
|
59 |
+
use_auth_token=os.environ['Read']
|
60 |
+
)
|
61 |
+
|
62 |
+
unet = UNet2DConditionModel.from_pretrained(
|
63 |
+
"NightRaven109/CCSRModels",
|
64 |
+
subfolder="stable-diffusion-2-1-base/unet",
|
65 |
+
use_auth_token=os.environ['Read']
|
66 |
+
)
|
67 |
+
|
68 |
+
controlnet = ControlNetModel.from_pretrained(
|
69 |
+
"NightRaven109/CCSRModels",
|
70 |
+
subfolder="Controlnet",
|
71 |
+
use_auth_token=os.environ['Read']
|
72 |
+
)
|
73 |
+
|
74 |
+
vae = AutoencoderKL.from_pretrained(
|
75 |
+
"NightRaven109/CCSRModels",
|
76 |
+
subfolder="vae",
|
77 |
+
use_auth_token=os.environ['Read']
|
78 |
+
)
|
79 |
+
|
80 |
+
# Rest of the code remains the same
|
81 |
+
# Freeze models
|
82 |
+
for model in [vae, text_encoder, unet, controlnet]:
|
83 |
+
model.requires_grad_(False)
|
84 |
+
|
85 |
+
# Initialize pipeline
|
86 |
+
pipeline = StableDiffusionControlNetPipeline(
|
87 |
+
vae=vae,
|
88 |
+
text_encoder=text_encoder,
|
89 |
+
tokenizer=tokenizer,
|
90 |
+
feature_extractor=feature_extractor,
|
91 |
+
unet=unet,
|
92 |
+
controlnet=controlnet,
|
93 |
+
scheduler=scheduler,
|
94 |
+
safety_checker=None,
|
95 |
+
requires_safety_checker=False,
|
96 |
+
)
|
97 |
+
|
98 |
+
# Get weight dtype based on mixed precision
|
99 |
+
weight_dtype = torch.float32
|
100 |
+
if accelerator.mixed_precision == "fp16":
|
101 |
+
weight_dtype = torch.float16
|
102 |
+
elif accelerator.mixed_precision == "bf16":
|
103 |
+
weight_dtype = torch.bfloat16
|
104 |
+
|
105 |
+
# Move models to device with appropriate dtype
|
106 |
+
for model in [text_encoder, vae, unet, controlnet]:
|
107 |
+
model.to(accelerator.device, dtype=weight_dtype)
|
108 |
+
|
109 |
+
# Initialize generator
|
110 |
+
generator = torch.Generator(device=accelerator.device)
|
111 |
+
|
112 |
+
return True
|
113 |
+
|
114 |
+
except Exception as e:
|
115 |
+
print(f"Error initializing models: {str(e)}")
|
116 |
+
return False
|
117 |
+
|
118 |
+
@spaces.GPU
|
119 |
+
def process_image(
|
120 |
+
input_image,
|
121 |
+
prompt="clean, high-resolution, 8k",
|
122 |
+
negative_prompt="blurry, dotted, noise, raster lines, unclear, lowres, over-smoothed",
|
123 |
+
guidance_scale=1.0,
|
124 |
+
conditioning_scale=1.0,
|
125 |
+
num_inference_steps=20,
|
126 |
+
seed=42,
|
127 |
+
upscale_factor=2,
|
128 |
+
color_fix_method="adain"
|
129 |
+
):
|
130 |
+
global pipeline, generator, accelerator
|
131 |
+
|
132 |
+
if pipeline is None:
|
133 |
+
if not initialize_models():
|
134 |
+
return None
|
135 |
+
|
136 |
+
try:
|
137 |
+
# Set seed
|
138 |
+
if seed is not None:
|
139 |
+
generator.manual_seed(seed)
|
140 |
+
|
141 |
+
# Process input image
|
142 |
+
input_pil = Image.fromarray(input_image)
|
143 |
+
width, height = input_pil.size
|
144 |
+
|
145 |
+
# Resize image
|
146 |
+
target_width = width * upscale_factor
|
147 |
+
target_height = height * upscale_factor
|
148 |
+
target_width = target_width - (target_width % 8)
|
149 |
+
target_height = target_height - (target_height % 8)
|
150 |
+
|
151 |
+
# Move pipeline to GPU for processing
|
152 |
+
pipeline.to(accelerator.device)
|
153 |
+
|
154 |
+
# Generate image
|
155 |
+
with torch.no_grad():
|
156 |
+
output = pipeline(
|
157 |
+
t_max=0.6666,
|
158 |
+
t_min=0.0,
|
159 |
+
tile_diffusion=False,
|
160 |
+
added_prompt=prompt,
|
161 |
+
image=input_pil,
|
162 |
+
num_inference_steps=num_inference_steps,
|
163 |
+
generator=generator,
|
164 |
+
height=target_height,
|
165 |
+
width=target_width,
|
166 |
+
guidance_scale=guidance_scale,
|
167 |
+
negative_prompt=negative_prompt,
|
168 |
+
conditioning_scale=conditioning_scale,
|
169 |
+
)
|
170 |
+
|
171 |
+
generated_image = output.images[0]
|
172 |
+
|
173 |
+
# Apply color fixing if specified
|
174 |
+
if color_fix_method != "none":
|
175 |
+
fix_func = wavelet_color_fix if color_fix_method == "wavelet" else adain_color_fix
|
176 |
+
generated_image = fix_func(generated_image, input_pil)
|
177 |
+
|
178 |
+
# Move pipeline back to CPU
|
179 |
+
pipeline.to("cpu")
|
180 |
+
torch.cuda.empty_cache()
|
181 |
+
|
182 |
+
return generated_image
|
183 |
+
|
184 |
+
except Exception as e:
|
185 |
+
print(f"Error processing image: {str(e)}")
|
186 |
+
return None
|
187 |
+
|
188 |
+
# Create Gradio interface
|
189 |
+
iface = gr.Interface(
|
190 |
+
fn=process_image,
|
191 |
+
inputs=[
|
192 |
+
gr.Image(label="Input Image"),
|
193 |
+
gr.Textbox(label="Prompt", value="clean, high-resolution, 8k"),
|
194 |
+
gr.Textbox(label="Negative Prompt", value="blurry, dotted, noise, raster lines, unclear, lowres, over-smoothed"),
|
195 |
+
gr.Slider(minimum=1.0, maximum=20.0, value=1.0, label="Guidance Scale"),
|
196 |
+
gr.Slider(minimum=0.1, maximum=2.0, value=1.0, label="Conditioning Scale"),
|
197 |
+
gr.Slider(minimum=1, maximum=50, value=20, step=1, label="Number of Steps"),
|
198 |
+
gr.Number(label="Seed", value=42),
|
199 |
+
gr.Slider(minimum=1, maximum=4, value=2, step=1, label="Upscale Factor"),
|
200 |
+
gr.Radio(["none", "wavelet", "adain"], label="Color Fix Method", value="adain"),
|
201 |
+
],
|
202 |
+
outputs=gr.Image(label="Generated Image"),
|
203 |
+
title="Controllable Conditional Super-Resolution",
|
204 |
+
description="Upload an image to enhance its resolution using CCSR.",
|
205 |
+
examples=[
|
206 |
+
["example1.jpg", "clean, sharp, detailed", "blurry, noise", 1.0, 1.0, 20, 42, 2, "adain"],
|
207 |
+
["example2.jpg", "high-resolution, pristine", "artifacts, pixelated", 1.5, 1.0, 30, 123, 2, "wavelet"],
|
208 |
+
]
|
209 |
+
)
|
210 |
+
|
211 |
+
if __name__ == "__main__":
|
212 |
+
iface.launch()
|
requirements.txt
CHANGED
@@ -1,14 +1,15 @@
|
|
1 |
-
diffusers==0.21.0
|
2 |
-
torch==2.0.1
|
3 |
-
pytorch_lightning
|
4 |
-
accelerate==1.2.0
|
5 |
-
transformers==4.25.0
|
6 |
-
xformers==0.0.22
|
7 |
-
loralib
|
8 |
-
fairscale==0.4.13
|
9 |
-
basicsr==1.4.2
|
10 |
-
timm==0.9.5
|
11 |
-
pydantic==1.10.11
|
12 |
-
huggingface_hub==0.25.2
|
13 |
-
opencv-python-headless
|
14 |
-
lpips
|
|
|
|
1 |
+
diffusers==0.21.0
|
2 |
+
torch==2.0.1
|
3 |
+
pytorch_lightning
|
4 |
+
accelerate==1.2.0
|
5 |
+
transformers==4.25.0
|
6 |
+
xformers==0.0.22
|
7 |
+
loralib
|
8 |
+
fairscale==0.4.13
|
9 |
+
basicsr==1.4.2
|
10 |
+
timm==0.9.5
|
11 |
+
pydantic==1.10.11
|
12 |
+
huggingface_hub==0.25.2
|
13 |
+
opencv-python-headless
|
14 |
+
lpips
|
15 |
+
einops
|