Spaces:
Running
on
Zero
Running
on
Zero
Upload 2 files
Browse files- app.py +302 -0
- requirements.txt +16 -0
app.py
ADDED
@@ -0,0 +1,302 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys
|
2 |
+
sys.path.append('../')
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import random
|
6 |
+
import numpy as np
|
7 |
+
from PIL import Image
|
8 |
+
|
9 |
+
import gradio as gr
|
10 |
+
from huggingface_hub import hf_hub_download
|
11 |
+
from transformers import AutoModelForImageSegmentation
|
12 |
+
from torchvision import transforms
|
13 |
+
|
14 |
+
from pipeline import InstantCharacterFluxPipeline
|
15 |
+
|
16 |
+
# global variable
|
17 |
+
MAX_SEED = np.iinfo(np.int32).max
|
18 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
19 |
+
dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32
|
20 |
+
|
21 |
+
# pre-trained weights
|
22 |
+
ip_adapter_path = hf_hub_download(repo_id="InstantX/InstantCharacter", filename="instantcharacter_ip-adapter.bin")
|
23 |
+
base_model = 'black-forest-labs/FLUX.1-dev'
|
24 |
+
image_encoder_path = 'google/siglip-so400m-patch14-384'
|
25 |
+
image_encoder_2_path = 'facebook/dinov2-giant'
|
26 |
+
birefnet_path = 'ZhengPeng7/BiRefNet'
|
27 |
+
makoto_style_lora_path = hf_hub_download(repo_id="InstantX/FLUX.1-dev-LoRA-Makoto-Shinkai", filename="Makoto_Shinkai_style.safetensors")
|
28 |
+
ghibli_style_lora_path = hf_hub_download(repo_id="InstantX/FLUX.1-dev-LoRA-Ghibli", filename="ghibli_style.safetensors")
|
29 |
+
|
30 |
+
# init InstantCharacter pipeline
|
31 |
+
pipe = InstantCharacterFluxPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16)
|
32 |
+
pipe.to(device)
|
33 |
+
|
34 |
+
# load InstantCharacter
|
35 |
+
pipe.init_adapter(
|
36 |
+
image_encoder_path=image_encoder_path,
|
37 |
+
image_encoder_2_path=image_encoder_2_path,
|
38 |
+
subject_ipadapter_cfg=dict(subject_ip_adapter_path=ip_adapter_path, nb_token=1024),
|
39 |
+
)
|
40 |
+
|
41 |
+
# load matting model
|
42 |
+
birefnet = AutoModelForImageSegmentation.from_pretrained(birefnet_path, trust_remote_code=True)
|
43 |
+
birefnet.to('cuda')
|
44 |
+
birefnet.eval()
|
45 |
+
birefnet_transform_image = transforms.Compose([
|
46 |
+
transforms.Resize((1024, 1024)),
|
47 |
+
transforms.ToTensor(),
|
48 |
+
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
49 |
+
])
|
50 |
+
|
51 |
+
|
52 |
+
def remove_bkg(subject_image):
|
53 |
+
|
54 |
+
def infer_matting(img_pil):
|
55 |
+
input_images = birefnet_transform_image(img_pil).unsqueeze(0).to('cuda')
|
56 |
+
|
57 |
+
with torch.no_grad():
|
58 |
+
preds = birefnet(input_images)[-1].sigmoid().cpu()
|
59 |
+
pred = preds[0].squeeze()
|
60 |
+
pred_pil = transforms.ToPILImage()(pred)
|
61 |
+
mask = pred_pil.resize(img_pil.size)
|
62 |
+
mask = np.array(mask)
|
63 |
+
mask = mask[..., None]
|
64 |
+
return mask
|
65 |
+
|
66 |
+
def get_bbox_from_mask(mask, th=128):
|
67 |
+
height, width = mask.shape[:2]
|
68 |
+
x1, y1, x2, y2 = 0, 0, width - 1, height - 1
|
69 |
+
|
70 |
+
sample = np.max(mask, axis=0)
|
71 |
+
for idx in range(width):
|
72 |
+
if sample[idx] >= th:
|
73 |
+
x1 = idx
|
74 |
+
break
|
75 |
+
|
76 |
+
sample = np.max(mask[:, ::-1], axis=0)
|
77 |
+
for idx in range(width):
|
78 |
+
if sample[idx] >= th:
|
79 |
+
x2 = width - 1 - idx
|
80 |
+
break
|
81 |
+
|
82 |
+
sample = np.max(mask, axis=1)
|
83 |
+
for idx in range(height):
|
84 |
+
if sample[idx] >= th:
|
85 |
+
y1 = idx
|
86 |
+
break
|
87 |
+
|
88 |
+
sample = np.max(mask[::-1], axis=1)
|
89 |
+
for idx in range(height):
|
90 |
+
if sample[idx] >= th:
|
91 |
+
y2 = height - 1 - idx
|
92 |
+
break
|
93 |
+
|
94 |
+
x1 = np.clip(x1, 0, width-1).round().astype(np.int32)
|
95 |
+
y1 = np.clip(y1, 0, height-1).round().astype(np.int32)
|
96 |
+
x2 = np.clip(x2, 0, width-1).round().astype(np.int32)
|
97 |
+
y2 = np.clip(y2, 0, height-1).round().astype(np.int32)
|
98 |
+
|
99 |
+
return [x1, y1, x2, y2]
|
100 |
+
|
101 |
+
def pad_to_square(image, pad_value = 255, random = False):
|
102 |
+
'''
|
103 |
+
image: np.array [h, w, 3]
|
104 |
+
'''
|
105 |
+
H,W = image.shape[0], image.shape[1]
|
106 |
+
if H == W:
|
107 |
+
return image
|
108 |
+
|
109 |
+
padd = abs(H - W)
|
110 |
+
if random:
|
111 |
+
padd_1 = int(np.random.randint(0,padd))
|
112 |
+
else:
|
113 |
+
padd_1 = int(padd / 2)
|
114 |
+
padd_2 = padd - padd_1
|
115 |
+
|
116 |
+
if H > W:
|
117 |
+
pad_param = ((0,0),(padd_1,padd_2),(0,0))
|
118 |
+
else:
|
119 |
+
pad_param = ((padd_1,padd_2),(0,0),(0,0))
|
120 |
+
|
121 |
+
image = np.pad(image, pad_param, 'constant', constant_values=pad_value)
|
122 |
+
return image
|
123 |
+
|
124 |
+
salient_object_mask = infer_matting(subject_image)[..., 0]
|
125 |
+
x1, y1, x2, y2 = get_bbox_from_mask(salient_object_mask)
|
126 |
+
subject_image = np.array(subject_image)
|
127 |
+
salient_object_mask[salient_object_mask > 128] = 255
|
128 |
+
salient_object_mask[salient_object_mask < 128] = 0
|
129 |
+
sample_mask = np.concatenate([salient_object_mask[..., None]]*3, axis=2)
|
130 |
+
obj_image = sample_mask / 255 * subject_image + (1 - sample_mask / 255) * 255
|
131 |
+
crop_obj_image = obj_image[y1:y2, x1:x2]
|
132 |
+
crop_pad_obj_image = pad_to_square(crop_obj_image, 255)
|
133 |
+
subject_image = Image.fromarray(crop_pad_obj_image.astype(np.uint8))
|
134 |
+
return subject_image
|
135 |
+
|
136 |
+
|
137 |
+
def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
|
138 |
+
if randomize_seed:
|
139 |
+
seed = random.randint(0, MAX_SEED)
|
140 |
+
return seed
|
141 |
+
|
142 |
+
def get_example():
|
143 |
+
case = [
|
144 |
+
[
|
145 |
+
"./assets/girl.jpg",
|
146 |
+
"A girl is playing a guitar in street",
|
147 |
+
0.9,
|
148 |
+
'Makoto Shinkai style',
|
149 |
+
],
|
150 |
+
[
|
151 |
+
"./assets/boy.jpg",
|
152 |
+
"A boy is riding a bike in snow",
|
153 |
+
0.9,
|
154 |
+
'Makoto Shinkai style',
|
155 |
+
],
|
156 |
+
]
|
157 |
+
return case
|
158 |
+
|
159 |
+
def run_for_examples(source_image, prompt, scale, style_mode):
|
160 |
+
|
161 |
+
return create_image(
|
162 |
+
input_image=source_image,
|
163 |
+
prompt=prompt,
|
164 |
+
scale=scale,
|
165 |
+
guidance_scale=3.5,
|
166 |
+
num_inference_steps=28,
|
167 |
+
seed=123456,
|
168 |
+
style_mode=style_mode,
|
169 |
+
)
|
170 |
+
|
171 |
+
def create_image(input_image,
|
172 |
+
prompt,
|
173 |
+
scale,
|
174 |
+
guidance_scale,
|
175 |
+
num_inference_steps,
|
176 |
+
seed,
|
177 |
+
style_mode=None):
|
178 |
+
|
179 |
+
input_image = remove_bkg(input_image)
|
180 |
+
|
181 |
+
if style_mode is None:
|
182 |
+
images = pipe(
|
183 |
+
prompt=prompt,
|
184 |
+
num_inference_steps=num_inference_steps,
|
185 |
+
guidance_scale=guidance_scale,
|
186 |
+
width=1024,
|
187 |
+
height=1024,
|
188 |
+
subject_image=input_image,
|
189 |
+
subject_scale=scale,
|
190 |
+
generator=torch.manual_seed(seed),
|
191 |
+
).images
|
192 |
+
else:
|
193 |
+
if style_mode == 'Makoto Shinkai style':
|
194 |
+
lora_file_path = makoto_style_lora_path
|
195 |
+
trigger = 'Makoto Shinkai style'
|
196 |
+
elif style_mode == 'Ghibli style':
|
197 |
+
lora_file_path = ghibli_style_lora_path
|
198 |
+
trigger = 'ghibli style'
|
199 |
+
|
200 |
+
images = pipe.with_style_lora(
|
201 |
+
lora_file_path=lora_file_path,
|
202 |
+
trigger=trigger,
|
203 |
+
prompt=prompt,
|
204 |
+
num_inference_steps=num_inference_steps,
|
205 |
+
guidance_scale=guidance_scale,
|
206 |
+
width=1024,
|
207 |
+
height=1024,
|
208 |
+
subject_image=input_image,
|
209 |
+
subject_scale=scale,
|
210 |
+
generator=torch.manual_seed(seed),
|
211 |
+
).images
|
212 |
+
|
213 |
+
|
214 |
+
return images
|
215 |
+
|
216 |
+
# Description
|
217 |
+
title = r"""
|
218 |
+
<h1 align="center">InstantCharacter : Personalize Any Characters with a Scalable Diffusion Transformer Framework</h1>
|
219 |
+
"""
|
220 |
+
|
221 |
+
description = r"""
|
222 |
+
<b>Official π€ Gradio demo</b> for <a href='https://instantcharacter.github.io/' target='_blank'><b>InstantCharacter : Personalize Any Characters with a Scalable Diffusion Transformer Framework</b></a>.<br>
|
223 |
+
How to use:<br>
|
224 |
+
1. Upload a character image, removing background would be preferred.
|
225 |
+
2. Enter a text prompt to describe what you hope the chracter does.
|
226 |
+
3. Click the <b>Submit</b> button to begin customization.
|
227 |
+
4. Share your custimized photo with your friends and enjoy! π
|
228 |
+
"""
|
229 |
+
|
230 |
+
article = r"""
|
231 |
+
---
|
232 |
+
π **Citation**
|
233 |
+
<br>
|
234 |
+
If our work is helpful for your research or applications, please cite us via:
|
235 |
+
```bibtex
|
236 |
+
TBD
|
237 |
+
```
|
238 |
+
π§ **Contact**
|
239 |
+
<br>
|
240 |
+
If you have any questions, please feel free to open an issue.
|
241 |
+
"""
|
242 |
+
|
243 |
+
block = gr.Blocks(css="footer {visibility: hidden}").queue(max_size=10, api_open=False)
|
244 |
+
with block:
|
245 |
+
|
246 |
+
# description
|
247 |
+
gr.Markdown(title)
|
248 |
+
gr.Markdown(description)
|
249 |
+
|
250 |
+
with gr.Tabs():
|
251 |
+
with gr.Row():
|
252 |
+
with gr.Column():
|
253 |
+
|
254 |
+
with gr.Row():
|
255 |
+
with gr.Column():
|
256 |
+
image_pil = gr.Image(label="Source Image", type='pil')
|
257 |
+
|
258 |
+
prompt = gr.Textbox(label="Prompt", value="a character is riding a bike in snow")
|
259 |
+
|
260 |
+
scale = gr.Slider(minimum=0, maximum=1.5, step=0.01,value=1.0, label="Scale")
|
261 |
+
style_mode = gr.Dropdown(label='Style', choices=[None, 'Makoto Shinkai style', 'Ghibli style'], value=None)
|
262 |
+
|
263 |
+
with gr.Accordion(open=False, label="Advanced Options"):
|
264 |
+
guidance_scale = gr.Slider(minimum=1,maximum=7.0, step=0.01,value=3.5, label="guidance scale")
|
265 |
+
num_inference_steps = gr.Slider(minimum=5,maximum=50.0, step=1.0,value=28, label="num inference steps")
|
266 |
+
seed = gr.Slider(minimum=-1000000, maximum=1000000, value=123456, step=1, label="Seed Value")
|
267 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
268 |
+
|
269 |
+
generate_button = gr.Button("Generate Image")
|
270 |
+
|
271 |
+
with gr.Column():
|
272 |
+
generated_image = gr.Gallery(label="Generated Image")
|
273 |
+
|
274 |
+
generate_button.click(
|
275 |
+
fn=randomize_seed_fn,
|
276 |
+
inputs=[seed, randomize_seed],
|
277 |
+
outputs=seed,
|
278 |
+
queue=False,
|
279 |
+
api_name=False,
|
280 |
+
).then(
|
281 |
+
fn=create_image,
|
282 |
+
inputs=[image_pil,
|
283 |
+
prompt,
|
284 |
+
scale,
|
285 |
+
guidance_scale,
|
286 |
+
num_inference_steps,
|
287 |
+
seed,
|
288 |
+
style_mode,
|
289 |
+
],
|
290 |
+
outputs=[generated_image])
|
291 |
+
|
292 |
+
gr.Examples(
|
293 |
+
examples=get_example(),
|
294 |
+
inputs=[image_pil, prompt, scale, style_mode],
|
295 |
+
fn=run_for_examples,
|
296 |
+
outputs=[generated_image],
|
297 |
+
cache_examples=True,
|
298 |
+
)
|
299 |
+
|
300 |
+
gr.Markdown(article)
|
301 |
+
|
302 |
+
block.launch(server_name="0.0.0.0", server_port=80)
|
requirements.txt
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
diffusers>=0.32.2
|
2 |
+
torch>=2.0.0
|
3 |
+
torchvision>=0.15.1
|
4 |
+
transformers>=4.37.1
|
5 |
+
accelerate
|
6 |
+
safetensors
|
7 |
+
einops
|
8 |
+
spaces>=0.19.4
|
9 |
+
omegaconf
|
10 |
+
peft
|
11 |
+
huggingface-hub>=0.20.2
|
12 |
+
opencv-python
|
13 |
+
gradio
|
14 |
+
controlnet_aux
|
15 |
+
gdown
|
16 |
+
peft
|