Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,144 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import einops
|
3 |
+
import gradio as gr
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import random
|
7 |
+
from PIL import Image
|
8 |
+
from pathlib import Path
|
9 |
+
from torchvision import transforms
|
10 |
+
import torch.nn.functional as F
|
11 |
+
from torchvision.models import resnet50, ResNet50_Weights
|
12 |
+
|
13 |
+
from pytorch_lightning import seed_everything
|
14 |
+
from transformers import CLIPTextModel, CLIPTokenizer, CLIPImageProcessor
|
15 |
+
from diffusers import AutoencoderKL, DDIMScheduler, PNDMScheduler, DPMSolverMultistepScheduler, UniPCMultistepScheduler
|
16 |
+
|
17 |
+
from pipelines.pipeline_pasd import StableDiffusionControlNetPipeline
|
18 |
+
from myutils.misc import load_dreambooth_lora, rand_name
|
19 |
+
from myutils.wavelet_color_fix import wavelet_color_fix
|
20 |
+
from annotator.retinaface import RetinaFaceDetection
|
21 |
+
|
22 |
+
use_pasd_light = False
|
23 |
+
face_detector = RetinaFaceDetection()
|
24 |
+
|
25 |
+
if use_pasd_light:
|
26 |
+
from models.pasd_light.unet_2d_condition import UNet2DConditionModel
|
27 |
+
from models.pasd_light.controlnet import ControlNetModel
|
28 |
+
else:
|
29 |
+
from models.pasd.unet_2d_condition import UNet2DConditionModel
|
30 |
+
from models.pasd.controlnet import ControlNetModel
|
31 |
+
|
32 |
+
pretrained_model_path = "checkpoints/stable-diffusion-v1-5"
|
33 |
+
ckpt_path = "runs/pasd/checkpoint-100000"
|
34 |
+
#dreambooth_lora_path = "checkpoints/personalized_models/toonyou_beta3.safetensors"
|
35 |
+
dreambooth_lora_path = "checkpoints/personalized_models/majicmixRealistic_v6.safetensors"
|
36 |
+
#dreambooth_lora_path = "checkpoints/personalized_models/Realistic_Vision_V5.1.safetensors"
|
37 |
+
weight_dtype = torch.float16
|
38 |
+
device = "cuda"
|
39 |
+
|
40 |
+
scheduler = UniPCMultistepScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
|
41 |
+
text_encoder = CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder")
|
42 |
+
tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")
|
43 |
+
vae = AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae")
|
44 |
+
feature_extractor = CLIPImageProcessor.from_pretrained(f"{pretrained_model_path}/feature_extractor")
|
45 |
+
unet = UNet2DConditionModel.from_pretrained(ckpt_path, subfolder="unet")
|
46 |
+
controlnet = ControlNetModel.from_pretrained(ckpt_path, subfolder="controlnet")
|
47 |
+
vae.requires_grad_(False)
|
48 |
+
text_encoder.requires_grad_(False)
|
49 |
+
unet.requires_grad_(False)
|
50 |
+
controlnet.requires_grad_(False)
|
51 |
+
|
52 |
+
unet, vae, text_encoder = load_dreambooth_lora(unet, vae, text_encoder, dreambooth_lora_path)
|
53 |
+
|
54 |
+
text_encoder.to(device, dtype=weight_dtype)
|
55 |
+
vae.to(device, dtype=weight_dtype)
|
56 |
+
unet.to(device, dtype=weight_dtype)
|
57 |
+
controlnet.to(device, dtype=weight_dtype)
|
58 |
+
|
59 |
+
validation_pipeline = StableDiffusionControlNetPipeline(
|
60 |
+
vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, feature_extractor=feature_extractor,
|
61 |
+
unet=unet, controlnet=controlnet, scheduler=scheduler, safety_checker=None, requires_safety_checker=False,
|
62 |
+
)
|
63 |
+
#validation_pipeline.enable_vae_tiling()
|
64 |
+
validation_pipeline._init_tiled_vae(decoder_tile_size=224)
|
65 |
+
|
66 |
+
weights = ResNet50_Weights.DEFAULT
|
67 |
+
preprocess = weights.transforms()
|
68 |
+
resnet = resnet50(weights=weights)
|
69 |
+
resnet.eval()
|
70 |
+
|
71 |
+
def inference(input_image, prompt, a_prompt, n_prompt, denoise_steps, upscale, alpha, cfg, seed):
|
72 |
+
process_size = 768
|
73 |
+
resize_preproc = transforms.Compose([
|
74 |
+
transforms.Resize(process_size, interpolation=transforms.InterpolationMode.BILINEAR),
|
75 |
+
])
|
76 |
+
|
77 |
+
with torch.no_grad():
|
78 |
+
seed_everything(seed)
|
79 |
+
generator = torch.Generator(device=device)
|
80 |
+
|
81 |
+
input_image = input_image.convert('RGB')
|
82 |
+
batch = preprocess(input_image).unsqueeze(0)
|
83 |
+
prediction = resnet(batch).squeeze(0).softmax(0)
|
84 |
+
class_id = prediction.argmax().item()
|
85 |
+
score = prediction[class_id].item()
|
86 |
+
category_name = weights.meta["categories"][class_id]
|
87 |
+
if score >= 0.1:
|
88 |
+
prompt += f"{category_name}" if prompt=='' else f", {category_name}"
|
89 |
+
|
90 |
+
prompt = a_prompt if prompt=='' else f"{prompt}, {a_prompt}"
|
91 |
+
|
92 |
+
ori_width, ori_height = input_image.size
|
93 |
+
resize_flag = False
|
94 |
+
|
95 |
+
rscale = upscale
|
96 |
+
input_image = input_image.resize((input_image.size[0]*rscale, input_image.size[1]*rscale))
|
97 |
+
|
98 |
+
if min(validation_image.size) < process_size:
|
99 |
+
validation_image = resize_preproc(validation_image)
|
100 |
+
|
101 |
+
input_image = input_image.resize((input_image.size[0]//8*8, input_image.size[1]//8*8))
|
102 |
+
width, height = input_image.size
|
103 |
+
resize_flag = True #
|
104 |
+
|
105 |
+
try:
|
106 |
+
image = validation_pipeline(
|
107 |
+
None, prompt, input_image, num_inference_steps=denoise_steps, generator=generator, height=height, width=width, guidance_scale=cfg,
|
108 |
+
negative_prompt=n_prompt, conditioning_scale=alpha, eta=0.0,
|
109 |
+
).images[0]
|
110 |
+
|
111 |
+
if True: #alpha<1.0:
|
112 |
+
image = wavelet_color_fix(image, input_image)
|
113 |
+
|
114 |
+
if resize_flag:
|
115 |
+
image = image.resize((ori_width*rscale, ori_height*rscale))
|
116 |
+
except Exception as e:
|
117 |
+
print(e)
|
118 |
+
image = Image.new(mode="RGB", size=(512, 512))
|
119 |
+
|
120 |
+
return image
|
121 |
+
|
122 |
+
title = "Pixel-Aware Stable Diffusion for Real-ISR"
|
123 |
+
description = "Gradio Demo for PASD Real-ISR. To use it, simply upload your image, or click one of the examples to load them."
|
124 |
+
article = "<p style='text-align: center'><a href='https://github.com/yangxy/PASD' target='_blank'>Github Repo Pytorch</a></p>"
|
125 |
+
examples=[['samples/27d38eeb2dbbe7c9.png'],['samples/629e4da70703193b.png']]
|
126 |
+
|
127 |
+
demo = gr.Interface(
|
128 |
+
fn=inference,
|
129 |
+
inputs=[gr.Image(type="pil"),
|
130 |
+
gr.Textbox(label="Prompt", value="Asian"),
|
131 |
+
gr.Textbox(label="Added Prompt", value='clean, high-resolution, 8k, best quality, masterpiece'),
|
132 |
+
gr.Textbox(label="Negative Prompt",value='dotted, noise, blur, lowres, oversmooth, longbody, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality'),
|
133 |
+
gr.Slider(label="Denoise Steps", minimum=10, maximum=50, value=20, step=1),
|
134 |
+
gr.Slider(label="Upsample Scale", minimum=1, maximum=4, value=2, step=1),
|
135 |
+
gr.Slider(label="Conditioning Scale", minimum=0.5, maximum=1.5, value=1.1, step=0.1),
|
136 |
+
gr.Slider(label="Classier-free Guidance", minimum=0.1, maximum=10.0, value=7.5, step=0.1),
|
137 |
+
gr.Slider(label="Seed", minimum=-1, maximum=2147483647, step=1, randomize=True)],
|
138 |
+
outputs=gr.Image(type="pil"),
|
139 |
+
title=title,
|
140 |
+
description=description,
|
141 |
+
article=article,
|
142 |
+
examples=examples).queue(concurrency_count=1)
|
143 |
+
|
144 |
+
demo.launch()
|