Spaces:
Paused
Paused
ldm version
Browse files- app_ldm.py +164 -0
app_ldm.py
ADDED
@@ -0,0 +1,164 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from PIL import Image
|
3 |
+
import json
|
4 |
+
import random
|
5 |
+
|
6 |
+
import cv2
|
7 |
+
import einops
|
8 |
+
import gradio as gr
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
|
12 |
+
from pytorch_lightning import seed_everything
|
13 |
+
from annotator.util import resize_image, HWC3
|
14 |
+
from cldm.model import create_model, load_state_dict
|
15 |
+
from cldm.ddim_hacked import DDIMSampler
|
16 |
+
|
17 |
+
import torch.nn as nn
|
18 |
+
from torch.nn.functional import threshold, normalize,interpolate
|
19 |
+
from torch.utils.data import Dataset
|
20 |
+
from torch.optim import Adam
|
21 |
+
from torch.utils.data import Dataset
|
22 |
+
from torchvision import transforms
|
23 |
+
from torch.utils.data import DataLoader
|
24 |
+
from transformers import SegformerFeatureExtractor, SegformerForSemanticSegmentation
|
25 |
+
|
26 |
+
import argparse
|
27 |
+
|
28 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
29 |
+
|
30 |
+
parseargs = argparse.ArgumentParser()
|
31 |
+
parseargs.add_argument('--model', type=str, default='control_sd15_colorize_epoch=156.ckpt')
|
32 |
+
args = parseargs.parse_args()
|
33 |
+
model_path = args.model
|
34 |
+
|
35 |
+
feature_extractor = SegformerFeatureExtractor.from_pretrained("matei-dorian/segformer-b5-finetuned-human-parsing")
|
36 |
+
segmodel = SegformerForSemanticSegmentation.from_pretrained("matei-dorian/segformer-b5-finetuned-human-parsing")
|
37 |
+
|
38 |
+
model = create_model('./models/control_sd15_colorize.yaml').cpu()
|
39 |
+
model.load_state_dict(load_state_dict(f"./models/{model_path}", location=device))
|
40 |
+
model = model.to(device)
|
41 |
+
ddim_sampler = DDIMSampler(model)
|
42 |
+
|
43 |
+
def LGB_TO_RGB(gray_image, rgb_image):
|
44 |
+
# gray_image [H, W, 1]
|
45 |
+
# rgb_image [H, W, 3]
|
46 |
+
|
47 |
+
lab_image = cv2.cvtColor(rgb_image, cv2.COLOR_RGB2LAB)
|
48 |
+
lab_image[:, :, 0] = gray_image[:, :, 0]
|
49 |
+
|
50 |
+
return cv2.cvtColor(lab_image, cv2.COLOR_LAB2RGB)
|
51 |
+
|
52 |
+
|
53 |
+
def process(input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, ddim_steps, guess_mode, strength, scale, seed, eta, threshold, save_memory=False):
|
54 |
+
# center crop image to square
|
55 |
+
# H, W, _ = input_image.shape
|
56 |
+
# if H > W:
|
57 |
+
# input_image = input_image[(H - W) // 2:(H + W) // 2, :, :]
|
58 |
+
# elif W > H:
|
59 |
+
# input_image = input_image[:, (W - H) // 2:(H + W) // 2, :]
|
60 |
+
|
61 |
+
with torch.no_grad():
|
62 |
+
img = resize_image(input_image, image_resolution)
|
63 |
+
H, W, C = img.shape
|
64 |
+
print("img shape: ", img.shape)
|
65 |
+
if C == 3:
|
66 |
+
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
|
67 |
+
detected_map = img[:, :, None]
|
68 |
+
print("Gray image shape: ", detected_map.shape)
|
69 |
+
control = torch.from_numpy(detected_map.copy()).float().to(device)
|
70 |
+
# control = einops.rearrange(control, 'h w c -> 1 c h w')
|
71 |
+
print("Control shape: ", control.shape)
|
72 |
+
|
73 |
+
control = control / 255.0
|
74 |
+
control = torch.stack([control for _ in range(num_samples)], dim=0)
|
75 |
+
print("Stacked control shape: ", control.shape)
|
76 |
+
control = einops.rearrange(control, 'b h w c -> b c h w').clone()
|
77 |
+
|
78 |
+
if seed == -1:
|
79 |
+
seed = random.randint(0, 65535)
|
80 |
+
seed_everything(seed)
|
81 |
+
|
82 |
+
if save_memory:
|
83 |
+
model.low_vram_shift(is_diffusing=False)
|
84 |
+
|
85 |
+
cond = {"c_concat": [control], "c_crossattn": [model.get_learned_conditioning([prompt + ', ' + a_prompt] * num_samples)]}
|
86 |
+
un_cond = {"c_concat": None if guess_mode else [control], "c_crossattn": [model.get_learned_conditioning([n_prompt] * num_samples)]}
|
87 |
+
shape = (4, H // 8, W // 8)
|
88 |
+
|
89 |
+
if save_memory:
|
90 |
+
model.low_vram_shift(is_diffusing=True)
|
91 |
+
|
92 |
+
model.control_scales = [strength * (0.825 ** float(12 - i)) for i in range(13)] if guess_mode else ([strength] * 13) # Magic number. IDK why. Perhaps because 0.825**12<0.01 but 0.826**12>0.01
|
93 |
+
samples, intermediates = ddim_sampler.sample(ddim_steps, num_samples,
|
94 |
+
shape, cond, verbose=False, eta=eta,
|
95 |
+
unconditional_guidance_scale=scale,
|
96 |
+
unconditional_conditioning=un_cond)
|
97 |
+
|
98 |
+
if save_memory:
|
99 |
+
model.low_vram_shift(is_diffusing=False)
|
100 |
+
|
101 |
+
x_samples = model.decode_first_stage(samples)
|
102 |
+
x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c') * 127.5 + 127.5).cpu().numpy().clip(0, 255).astype(np.uint8)
|
103 |
+
|
104 |
+
results = [x_samples[i] for i in range(num_samples)]
|
105 |
+
results = [LGB_TO_RGB(detected_map, result) for result in results]
|
106 |
+
|
107 |
+
# results의 각 이미지를 mask로 변환
|
108 |
+
masks = []
|
109 |
+
for result in results:
|
110 |
+
inputs = feature_extractor(images=result, return_tensors="pt")
|
111 |
+
outputs = segmodel(**inputs)
|
112 |
+
logits = outputs.logits
|
113 |
+
logits = logits.squeeze(0)
|
114 |
+
thresholded = torch.zeros_like(logits)
|
115 |
+
thresholded[logits > threshold] = 1
|
116 |
+
mask = thresholded[1: ,:, :].sum(dim=0)
|
117 |
+
mask = mask.unsqueeze(0).unsqueeze(0)
|
118 |
+
mask = interpolate(mask, size=(H, W), mode='bilinear')
|
119 |
+
mask = mask.detach().numpy()
|
120 |
+
mask = np.squeeze(mask)
|
121 |
+
mask = np.where(mask > threshold, 1, 0)
|
122 |
+
masks.append(mask)
|
123 |
+
|
124 |
+
# results의 각 이미지를 mask를 이용해 mask가 0인 부분은 img 즉 흑백 이미지로 변환.
|
125 |
+
# img를 channel이 3인 rgb 이미지로 변환
|
126 |
+
gray_img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) # [H, W, 3]
|
127 |
+
final = [gray_img * (1 - mask[:, :, None]) + result * mask[:, :, None] for result, mask in zip(results, masks)]
|
128 |
+
|
129 |
+
# mask to 255 img
|
130 |
+
|
131 |
+
mask_img = [mask * 255 for mask in masks]
|
132 |
+
return [detected_map.squeeze(-1)] + results + mask_img + final
|
133 |
+
|
134 |
+
|
135 |
+
block = gr.Blocks().queue()
|
136 |
+
with block:
|
137 |
+
with gr.Row():
|
138 |
+
gr.Markdown("## Control Stable Diffusion with Gray Image")
|
139 |
+
with gr.Row():
|
140 |
+
with gr.Column():
|
141 |
+
input_image = gr.Image(sources=['upload'], type="numpy")
|
142 |
+
prompt = gr.Textbox(label="Prompt")
|
143 |
+
run_button = gr.Button(value="Run")
|
144 |
+
with gr.Accordion("Advanced options", open=False):
|
145 |
+
num_samples = gr.Slider(label="Images", minimum=1, maximum=12, value=1, step=1)
|
146 |
+
image_resolution = gr.Slider(label="Image Resolution", minimum=256, maximum=768, value=512, step=64)
|
147 |
+
strength = gr.Slider(label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01)
|
148 |
+
guess_mode = gr.Checkbox(label='Guess Mode', value=False)
|
149 |
+
ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=20, step=1)
|
150 |
+
scale = gr.Slider(label="Guidance Scale", minimum=0.1, maximum=30.0, value=1.0, step=0.1)
|
151 |
+
threshold = gr.Slider(label="segmentation threshold", minimum=0.1, maximum=0.9, value=0.5, step=0.05)
|
152 |
+
seed = gr.Slider(label="Seed", minimum=-1, maximum=2147483647, step=1, randomize=True)
|
153 |
+
eta = gr.Number(label="eta (DDIM)", value=0.0)
|
154 |
+
a_prompt = gr.Textbox(label="Added Prompt", value='best quality, extremely detailed')
|
155 |
+
n_prompt = gr.Textbox(label="Negative Prompt",
|
156 |
+
value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality')
|
157 |
+
with gr.Column():
|
158 |
+
# result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto')
|
159 |
+
result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery")
|
160 |
+
ips = [input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, ddim_steps, guess_mode, strength, scale, seed, eta, threshold]
|
161 |
+
run_button.click(fn=process, inputs=ips, outputs=[result_gallery], concurrency_limit=4)
|
162 |
+
|
163 |
+
block.queue(max_size=100)
|
164 |
+
block.launch(share=True)
|