Spaces:
Running
on
Zero
Running
on
Zero
Apply formatter to app.py and requirements.txt
Browse files- app.py +397 -275
- requirements.txt +22 -23
app.py
CHANGED
@@ -1,35 +1,43 @@
|
|
1 |
import os
|
2 |
import sys
|
3 |
|
4 |
-
|
5 |
print("Installing correct gradio version...")
|
6 |
os.system("pip uninstall -y gradio")
|
7 |
os.system("pip install gradio==4.38.1")
|
8 |
print("Installing Finished!")
|
9 |
|
10 |
|
|
|
|
|
|
|
|
|
11 |
import gradio as gr
|
12 |
import numpy as np
|
13 |
-
import
|
14 |
-
import uuid
|
15 |
import torch
|
16 |
import torchvision
|
17 |
-
import
|
18 |
-
import spaces
|
19 |
-
|
20 |
-
from PIL import Image
|
21 |
-
from omegaconf import OmegaConf
|
22 |
from einops import rearrange, repeat
|
23 |
-
from
|
|
|
|
|
24 |
from transformers import CLIPTextModel, CLIPTokenizer
|
25 |
-
from diffusers import AutoencoderKL, DDIMScheduler
|
26 |
|
27 |
-
from pipelines.pipeline_imagecoductor import ImageConductorPipeline
|
28 |
from modules.unet import UNet3DConditionFlowModel
|
29 |
-
from
|
30 |
-
from utils.
|
31 |
from utils.lora_utils import add_LoRA_to_controlnet
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
from utils.visualizer import Visualizer, vis_flow_to_video
|
|
|
33 |
#### Description ####
|
34 |
title = r"""<h1 align="center">CustomNet: Object Customization with Variable-Viewpoints in Text-to-Image Diffusion Models</h1>"""
|
35 |
|
@@ -41,7 +49,7 @@ head = r"""
|
|
41 |
<a href='https://liyaowei-stu.github.io/project/ImageConductor/'><img src='https://img.shields.io/badge/Project_Page-ImgaeConductor-green' alt='Project Page'></a>
|
42 |
<a href='https://arxiv.org/pdf/2406.15339'><img src='https://img.shields.io/badge/Paper-Arxiv-blue'></a>
|
43 |
<a href='https://github.com/liyaowei-stu/ImageConductor'><img src='https://img.shields.io/badge/Code-Github-orange'></a>
|
44 |
-
|
45 |
|
46 |
</div>
|
47 |
</br>
|
@@ -49,7 +57,6 @@ head = r"""
|
|
49 |
"""
|
50 |
|
51 |
|
52 |
-
|
53 |
descriptions = r"""
|
54 |
Official Gradio Demo for <a href='https://github.com/liyaowei-stu/ImageConductor'><b>Image Conductor: Precision Control for Interactive Video Synthesis</b></a>.<br>
|
55 |
🧙Image Conductor enables precise, fine-grained control for generating motion-controllable videos from images, advancing the practical application of interactive video synthesis.<br>
|
@@ -66,7 +73,7 @@ instructions = r"""
|
|
66 |
"""
|
67 |
|
68 |
citation = r"""
|
69 |
-
If Image Conductor is helpful, please help to ⭐ the <a href='https://github.com/liyaowei-stu/ImageConductor' target='_blank'>Github Repo</a>. Thanks!
|
70 |
[![GitHub Stars](https://img.shields.io/github/stars/liyaowei-stu%2FImageConductor)](https://github.com/liyaowei-stu/ImageConductor)
|
71 |
---
|
72 |
|
@@ -75,7 +82,7 @@ If Image Conductor is helpful, please help to ⭐ the <a href='https://github.co
|
|
75 |
If our work is useful for your research, please consider citing:
|
76 |
```bibtex
|
77 |
@misc{li2024imageconductor,
|
78 |
-
title={Image Conductor: Precision Control for Interactive Video Synthesis},
|
79 |
author={Li, Yaowei and Wang, Xintao and Zhang, Zhaoyang and Wang, Zhouxia and Yuan, Ziyang and Xie, Liangbin and Zou, Yuexian and Shan, Ying},
|
80 |
year={2024},
|
81 |
eprint={2406.15339},
|
@@ -94,42 +101,68 @@ os.makedirs("models/personalized")
|
|
94 |
os.makedirs("models/sd1-5")
|
95 |
|
96 |
if not os.path.exists("models/flow_controlnet.ckpt"):
|
97 |
-
os.system(
|
98 |
-
|
99 |
-
|
|
|
|
|
|
|
|
|
100 |
|
101 |
if not os.path.exists("models/image_controlnet.ckpt"):
|
102 |
-
os.system(
|
103 |
-
|
104 |
-
|
|
|
|
|
|
|
|
|
105 |
|
106 |
if not os.path.exists("models/unet.ckpt"):
|
107 |
-
os.system(
|
108 |
-
|
109 |
-
|
|
|
|
|
|
|
|
|
|
|
110 |
|
111 |
-
|
112 |
if not os.path.exists("models/sd1-5/config.json"):
|
113 |
-
os.system(
|
114 |
-
|
115 |
-
|
|
|
|
|
|
|
|
|
116 |
|
117 |
|
118 |
if not os.path.exists("models/sd1-5/unet.ckpt"):
|
119 |
-
os.system(f
|
120 |
|
121 |
# os.system(f'wget https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/unet/diffusion_pytorch_model.bin?download=true -P models/sd1-5/')
|
122 |
|
123 |
if not os.path.exists("models/personalized/helloobjects_V12c.safetensors"):
|
124 |
-
os.system(
|
125 |
-
|
126 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
127 |
|
128 |
|
129 |
if not os.path.exists("models/personalized/TUSUN.safetensors"):
|
130 |
-
os.system(
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# mv1 = os.system(f'mv /usr/local/lib/python3.10/site-packages/gradio/helpers.py /usr/local/lib/python3.10/site-packages/gradio/helpers_bkp.py')
|
135 |
# mv2 = os.system(f'mv helpers.py /usr/local/lib/python3.10/site-packages/gradio/helpers.py')
|
@@ -145,128 +178,135 @@ if not os.path.exists("models/personalized/TUSUN.safetensors"):
|
|
145 |
# - - - - - examples - - - - - #
|
146 |
|
147 |
image_examples = [
|
148 |
-
[
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
[
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
[
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
]
|
203 |
|
204 |
|
205 |
POINTS = {
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
}
|
213 |
|
214 |
IMAGE_PATH = {
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
}
|
222 |
|
223 |
|
224 |
-
|
225 |
DREAM_BOOTH = {
|
226 |
-
|
227 |
}
|
228 |
|
229 |
LORA = {
|
230 |
-
|
231 |
}
|
232 |
|
233 |
LORA_ALPHA = {
|
234 |
-
|
235 |
}
|
236 |
|
237 |
NPROMPT = {
|
238 |
-
"HelloObject":
|
239 |
}
|
240 |
|
241 |
output_dir = "outputs"
|
242 |
ensure_dirname(output_dir)
|
243 |
|
|
|
244 |
def points_to_flows(track_points, model_length, height, width):
|
245 |
input_drag = np.zeros((model_length - 1, height, width, 2))
|
246 |
for splited_track in track_points:
|
247 |
-
if len(splited_track) == 1:
|
248 |
displacement_point = tuple([splited_track[0][0] + 1, splited_track[0][1] + 1])
|
249 |
splited_track = tuple([splited_track[0], displacement_point])
|
250 |
# interpolate the track
|
251 |
splited_track = interpolate_trajectory(splited_track, model_length)
|
252 |
splited_track = splited_track[:model_length]
|
253 |
if len(splited_track) < model_length:
|
254 |
-
splited_track = splited_track + [splited_track[-1]] * (model_length -len(splited_track))
|
255 |
for i in range(model_length - 1):
|
256 |
start_point = splited_track[i]
|
257 |
-
end_point = splited_track[i+1]
|
258 |
input_drag[i][int(start_point[1])][int(start_point[0])][0] = end_point[0] - start_point[0]
|
259 |
input_drag[i][int(start_point[1])][int(start_point[0])][1] = end_point[1] - start_point[1]
|
260 |
return input_drag
|
261 |
|
|
|
262 |
class ImageConductor:
|
263 |
-
def __init__(
|
|
|
|
|
264 |
self.device = device
|
265 |
-
tokenizer
|
266 |
-
text_encoder = CLIPTextModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="text_encoder").to(
|
267 |
-
|
|
|
|
|
268 |
inference_config = OmegaConf.load("configs/inference/inference.yaml")
|
269 |
-
unet = UNet3DConditionFlowModel.from_pretrained_2d(
|
|
|
|
|
270 |
|
271 |
self.vae = vae
|
272 |
|
@@ -287,15 +327,14 @@ class ImageConductor:
|
|
287 |
|
288 |
self.pipeline = ImageConductorPipeline(
|
289 |
unet=unet,
|
290 |
-
vae=vae,
|
291 |
-
tokenizer=tokenizer,
|
292 |
-
text_encoder=text_encoder,
|
293 |
scheduler=DDIMScheduler(**OmegaConf.to_container(inference_config.noise_scheduler_kwargs)),
|
294 |
image_controlnet=image_controlnet,
|
295 |
flow_controlnet=flow_controlnet,
|
296 |
).to(device)
|
297 |
|
298 |
-
|
299 |
self.height = height
|
300 |
self.width = width
|
301 |
# _, model_step, _ = split_filename(model_path)
|
@@ -307,7 +346,20 @@ class ImageConductor:
|
|
307 |
self.blur_kernel = blur_kernel
|
308 |
|
309 |
@spaces.GPU(duration=180)
|
310 |
-
def run(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
311 |
print("Run!")
|
312 |
if examples_type != "":
|
313 |
### for adapting high version gradio
|
@@ -317,30 +369,40 @@ class ImageConductor:
|
|
317 |
tracking_points.value.extend(points)
|
318 |
print("example first_frame_path", first_frame_path)
|
319 |
print("example tracking_points", tracking_points.value)
|
320 |
-
|
321 |
-
original_width, original_height=384, 256
|
322 |
if isinstance(tracking_points, list):
|
323 |
input_all_points = tracking_points
|
324 |
else:
|
325 |
input_all_points = tracking_points.value
|
326 |
-
|
327 |
print("input_all_points", input_all_points)
|
328 |
-
resized_all_points = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
329 |
|
330 |
dir, base, ext = split_filename(first_frame_path)
|
331 |
-
id = base.split(
|
332 |
-
|
333 |
-
|
334 |
-
|
|
|
335 |
|
336 |
-
## image condition
|
337 |
-
image_transforms = transforms.Compose(
|
|
|
338 |
transforms.RandomResizedCrop(
|
339 |
-
(self.height, self.width), (1.0, 1.0),
|
340 |
-
ratio=(self.width/self.height, self.width/self.height)
|
341 |
),
|
342 |
transforms.ToTensor(),
|
343 |
-
]
|
|
|
344 |
|
345 |
image_paths = [first_frame_path]
|
346 |
controlnet_images = [(image_transforms(Image.open(path).convert("RGB"))) for path in image_paths]
|
@@ -349,60 +411,68 @@ class ImageConductor:
|
|
349 |
num_controlnet_images = controlnet_images.shape[2]
|
350 |
controlnet_images = rearrange(controlnet_images, "b c f h w -> (b f) c h w")
|
351 |
self.vae.to(device)
|
352 |
-
controlnet_images = self.vae.encode(controlnet_images * 2. - 1.).latent_dist.sample() * 0.18215
|
353 |
controlnet_images = rearrange(controlnet_images, "(b f) c h w -> b c f h w", f=num_controlnet_images)
|
354 |
|
355 |
# flow condition
|
356 |
controlnet_flows = points_to_flows(resized_all_points, self.model_length, self.height, self.width)
|
357 |
-
for i in range(0, self.model_length-1):
|
358 |
controlnet_flows[i] = cv2.filter2D(controlnet_flows[i], -1, self.blur_kernel)
|
359 |
-
controlnet_flows = np.concatenate(
|
|
|
|
|
360 |
os.makedirs(os.path.join(output_dir, "control_flows"), exist_ok=True)
|
361 |
-
trajs_video = vis_flow_to_video(controlnet_flows, num_frames=self.model_length)
|
362 |
-
torchvision.io.write_video(
|
363 |
-
|
364 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
365 |
|
366 |
-
dreambooth_model_path = DREAM_BOOTH.get(personalized,
|
367 |
-
lora_model_path = LORA.get(personalized,
|
368 |
lora_alpha = LORA_ALPHA.get(personalized, 0.6)
|
369 |
self.pipeline = load_weights(
|
370 |
self.pipeline,
|
371 |
-
dreambooth_model_path
|
372 |
-
lora_model_path
|
373 |
-
lora_alpha
|
374 |
).to(device)
|
375 |
-
|
376 |
-
if NPROMPT.get(personalized,
|
377 |
-
negative_prompt =
|
378 |
-
|
379 |
if randomize_seed:
|
380 |
random_seed = torch.seed()
|
381 |
else:
|
382 |
seed = int(seed)
|
383 |
random_seed = seed
|
384 |
torch.manual_seed(random_seed)
|
385 |
-
torch.cuda.manual_seed_all(random_seed)
|
386 |
print(f"current seed: {torch.initial_seed()}")
|
387 |
sample = self.pipeline(
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
|
394 |
-
|
395 |
-
|
396 |
-
|
397 |
-
|
398 |
-
|
399 |
-
|
400 |
-
|
401 |
-
|
402 |
-
outputs_path = os.path.join(output_dir, f
|
403 |
-
vis_video = (rearrange(sample[0],
|
404 |
-
torchvision.io.write_video(outputs_path, vis_video, fps=8, video_codec=
|
405 |
-
|
406 |
# outputs_path = os.path.join(output_dir, f'output_{i}_{id}.gif')
|
407 |
# save_videos_grid(sample[0][None], outputs_path)
|
408 |
print("Done!")
|
@@ -412,33 +482,40 @@ class ImageConductor:
|
|
412 |
def reset_states(first_frame_path, tracking_points):
|
413 |
first_frame_path = gr.State()
|
414 |
tracking_points = gr.State([])
|
415 |
-
return {input_image:None, first_frame_path_var: first_frame_path, tracking_points_var: tracking_points}
|
416 |
|
417 |
|
418 |
def preprocess_image(image, tracking_points):
|
419 |
image_pil = image2pil(image.name)
|
420 |
raw_w, raw_h = image_pil.size
|
421 |
-
resize_ratio = max(384/raw_w, 256/raw_h)
|
422 |
image_pil = image_pil.resize((int(raw_w * resize_ratio), int(raw_h * resize_ratio)), Image.BILINEAR)
|
423 |
-
image_pil = transforms.CenterCrop((256, 384))(image_pil.convert(
|
424 |
id = str(uuid.uuid4())[:4]
|
425 |
first_frame_path = os.path.join(output_dir, f"first_frame_{id}.jpg")
|
426 |
image_pil.save(first_frame_path, quality=95)
|
427 |
-
tracking_points =
|
428 |
-
return {
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
433 |
color = (255, 0, 0, 255)
|
434 |
-
elif drag_mode==
|
435 |
color = (0, 0, 255, 255)
|
436 |
|
437 |
-
if not isinstance(tracking_points
|
438 |
print(f"You selected {evt.value} at {evt.index} from {evt.target}")
|
439 |
tracking_points.value[-1].append(evt.index)
|
440 |
print(tracking_points.value)
|
441 |
-
tracking_points_values =
|
442 |
else:
|
443 |
try:
|
444 |
tracking_points[-1].append(evt.index)
|
@@ -446,26 +523,33 @@ def add_tracking_points(tracking_points, first_frame_path, drag_mode, evt: gr.Se
|
|
446 |
tracking_points.append([])
|
447 |
tracking_points[-1].append(evt.index)
|
448 |
print(f"Solved Error: {e}")
|
449 |
-
|
450 |
tracking_points_values = tracking_points
|
451 |
-
|
452 |
-
|
453 |
-
transparent_background = Image.open(first_frame_path).convert('RGBA')
|
454 |
w, h = transparent_background.size
|
455 |
transparent_layer = np.zeros((h, w, 4))
|
456 |
-
|
457 |
for track in tracking_points_values:
|
458 |
if len(track) > 1:
|
459 |
-
for i in range(len(track)-1):
|
460 |
start_point = track[i]
|
461 |
-
end_point = track[i+1]
|
462 |
vx = end_point[0] - start_point[0]
|
463 |
vy = end_point[1] - start_point[1]
|
464 |
arrow_length = np.sqrt(vx**2 + vy**2)
|
465 |
-
if i == len(track)-2:
|
466 |
-
cv2.arrowedLine(
|
|
|
|
|
467 |
else:
|
468 |
-
cv2.line(
|
|
|
|
|
|
|
|
|
|
|
|
|
469 |
else:
|
470 |
cv2.circle(transparent_layer, tuple(track[0]), 5, color, -1)
|
471 |
|
@@ -475,79 +559,90 @@ def add_tracking_points(tracking_points, first_frame_path, drag_mode, evt: gr.Se
|
|
475 |
|
476 |
|
477 |
def add_drag(tracking_points):
|
478 |
-
if not isinstance(tracking_points
|
479 |
# print("before", tracking_points.value)
|
480 |
tracking_points.value.append([])
|
481 |
# print(tracking_points.value)
|
482 |
else:
|
483 |
tracking_points.append([])
|
484 |
return {tracking_points_var: tracking_points}
|
485 |
-
|
486 |
|
487 |
def delete_last_drag(tracking_points, first_frame_path, drag_mode):
|
488 |
-
if drag_mode==
|
489 |
color = (255, 0, 0, 255)
|
490 |
-
elif drag_mode==
|
491 |
color = (0, 0, 255, 255)
|
492 |
tracking_points.value.pop()
|
493 |
-
transparent_background = Image.open(first_frame_path).convert(
|
494 |
w, h = transparent_background.size
|
495 |
transparent_layer = np.zeros((h, w, 4))
|
496 |
for track in tracking_points.value:
|
497 |
if len(track) > 1:
|
498 |
-
for i in range(len(track)-1):
|
499 |
start_point = track[i]
|
500 |
-
end_point = track[i+1]
|
501 |
vx = end_point[0] - start_point[0]
|
502 |
vy = end_point[1] - start_point[1]
|
503 |
arrow_length = np.sqrt(vx**2 + vy**2)
|
504 |
-
if i == len(track)-2:
|
505 |
-
cv2.arrowedLine(
|
|
|
|
|
506 |
else:
|
507 |
-
cv2.line(
|
|
|
|
|
|
|
|
|
|
|
|
|
508 |
else:
|
509 |
cv2.circle(transparent_layer, tuple(track[0]), 5, color, -1)
|
510 |
|
511 |
transparent_layer = Image.fromarray(transparent_layer.astype(np.uint8))
|
512 |
trajectory_map = Image.alpha_composite(transparent_background, transparent_layer)
|
513 |
return {tracking_points_var: tracking_points, input_image: trajectory_map}
|
514 |
-
|
515 |
|
516 |
def delete_last_step(tracking_points, first_frame_path, drag_mode):
|
517 |
-
if drag_mode==
|
518 |
color = (255, 0, 0, 255)
|
519 |
-
elif drag_mode==
|
520 |
color = (0, 0, 255, 255)
|
521 |
tracking_points.value[-1].pop()
|
522 |
-
transparent_background = Image.open(first_frame_path).convert(
|
523 |
w, h = transparent_background.size
|
524 |
transparent_layer = np.zeros((h, w, 4))
|
525 |
for track in tracking_points.value:
|
526 |
if len(track) > 1:
|
527 |
-
for i in range(len(track)-1):
|
528 |
start_point = track[i]
|
529 |
-
end_point = track[i+1]
|
530 |
vx = end_point[0] - start_point[0]
|
531 |
vy = end_point[1] - start_point[1]
|
532 |
arrow_length = np.sqrt(vx**2 + vy**2)
|
533 |
-
if i == len(track)-2:
|
534 |
-
cv2.arrowedLine(
|
|
|
|
|
535 |
else:
|
536 |
-
cv2.line(
|
|
|
|
|
|
|
|
|
|
|
|
|
537 |
else:
|
538 |
-
cv2.circle(transparent_layer, tuple(track[0]), 5,color, -1)
|
539 |
|
540 |
transparent_layer = Image.fromarray(transparent_layer.astype(np.uint8))
|
541 |
trajectory_map = Image.alpha_composite(transparent_background, transparent_layer)
|
542 |
return {tracking_points_var: tracking_points, input_image: trajectory_map}
|
543 |
|
544 |
|
545 |
-
block = gr.Blocks(
|
546 |
-
theme=gr.themes.Soft(
|
547 |
-
radius_size=gr.themes.sizes.radius_none,
|
548 |
-
text_size=gr.themes.sizes.text_md
|
549 |
-
)
|
550 |
-
)
|
551 |
with block:
|
552 |
with gr.Row():
|
553 |
with gr.Column():
|
@@ -557,66 +652,72 @@ with block:
|
|
557 |
|
558 |
with gr.Accordion(label="🛠️ Instructions:", open=True, elem_id="accordion"):
|
559 |
with gr.Row(equal_height=True):
|
560 |
-
gr.Markdown(instructions)
|
561 |
-
|
562 |
|
563 |
# device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
564 |
device = torch.device("cuda")
|
565 |
-
unet_path =
|
566 |
-
image_controlnet_path =
|
567 |
-
flow_controlnet_path =
|
568 |
-
ImageConductor_net = ImageConductor(
|
569 |
-
|
570 |
-
|
571 |
-
|
572 |
-
|
573 |
-
|
574 |
-
|
575 |
-
|
|
|
576 |
first_frame_path_var = gr.State(value=None)
|
577 |
tracking_points_var = gr.State([])
|
578 |
|
579 |
with gr.Row():
|
580 |
with gr.Column(scale=1):
|
581 |
-
image_upload_button = gr.UploadButton(label="Upload Image",file_types=["image"])
|
582 |
add_drag_button = gr.Button(value="Add Drag")
|
583 |
reset_button = gr.Button(value="Reset")
|
584 |
delete_last_drag_button = gr.Button(value="Delete last drag")
|
585 |
delete_last_step_button = gr.Button(value="Delete last step")
|
586 |
-
|
587 |
-
|
588 |
|
589 |
with gr.Column(scale=7):
|
590 |
with gr.Row():
|
591 |
with gr.Column(scale=6):
|
592 |
-
input_image = gr.Image(
|
593 |
-
|
594 |
-
|
595 |
-
|
|
|
|
|
596 |
with gr.Column(scale=6):
|
597 |
-
output_image = gr.Image(
|
598 |
-
|
599 |
-
|
600 |
-
|
|
|
|
|
601 |
with gr.Row():
|
602 |
with gr.Column(scale=1):
|
603 |
-
prompt = gr.Textbox(
|
|
|
|
|
604 |
negative_prompt = gr.Text(
|
605 |
-
|
606 |
-
|
607 |
-
|
608 |
-
|
609 |
-
|
610 |
-
|
|
|
611 |
run_button = gr.Button(value="Run")
|
612 |
|
613 |
with gr.Accordion("More input params", open=False, elem_id="accordion1"):
|
614 |
with gr.Group():
|
615 |
seed = gr.Textbox(
|
616 |
-
label="Seed: ",
|
|
|
617 |
)
|
618 |
randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
|
619 |
-
|
620 |
with gr.Group():
|
621 |
with gr.Row():
|
622 |
guidance_scale = gr.Slider(
|
@@ -633,23 +734,18 @@ with block:
|
|
633 |
step=1,
|
634 |
value=25,
|
635 |
)
|
636 |
-
|
637 |
with gr.Group():
|
638 |
-
personalized = gr.Dropdown(label="Personalized", choices=["",
|
639 |
-
examples_type = gr.Textbox(label="Examples Type (Ignore) ",
|
640 |
|
641 |
with gr.Column(scale=7):
|
642 |
-
output_video = gr.Video(
|
643 |
-
label="Output Video",
|
644 |
-
width=384,
|
645 |
-
height=256)
|
646 |
# output_video = gr.Image(label="Output Video",
|
647 |
# height=256,
|
648 |
# width=384,)
|
649 |
-
|
650 |
-
|
651 |
with gr.Row():
|
652 |
-
|
653 |
|
654 |
example = gr.Examples(
|
655 |
label="Input Example",
|
@@ -658,26 +754,52 @@ with block:
|
|
658 |
examples_per_page=10,
|
659 |
cache_examples=False,
|
660 |
)
|
661 |
-
|
662 |
-
|
663 |
with gr.Row():
|
664 |
gr.Markdown(citation)
|
665 |
|
666 |
-
|
667 |
-
|
|
|
|
|
|
|
668 |
|
669 |
add_drag_button.click(add_drag, tracking_points_var, tracking_points_var)
|
670 |
|
671 |
-
delete_last_drag_button.click(
|
672 |
-
|
673 |
-
|
674 |
-
|
675 |
-
|
676 |
-
|
677 |
-
|
678 |
-
|
679 |
-
|
680 |
-
|
681 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
682 |
|
683 |
block.queue().launch()
|
|
|
1 |
import os
|
2 |
import sys
|
3 |
|
|
|
4 |
print("Installing correct gradio version...")
|
5 |
os.system("pip uninstall -y gradio")
|
6 |
os.system("pip install gradio==4.38.1")
|
7 |
print("Installing Finished!")
|
8 |
|
9 |
|
10 |
+
import json
|
11 |
+
import uuid
|
12 |
+
|
13 |
+
import cv2
|
14 |
import gradio as gr
|
15 |
import numpy as np
|
16 |
+
import spaces
|
|
|
17 |
import torch
|
18 |
import torchvision
|
19 |
+
from diffusers import AutoencoderKL, DDIMScheduler
|
|
|
|
|
|
|
|
|
20 |
from einops import rearrange, repeat
|
21 |
+
from omegaconf import OmegaConf
|
22 |
+
from PIL import Image
|
23 |
+
from torchvision import transforms, utils
|
24 |
from transformers import CLIPTextModel, CLIPTokenizer
|
|
|
25 |
|
|
|
26 |
from modules.unet import UNet3DConditionFlowModel
|
27 |
+
from pipelines.pipeline_imagecoductor import ImageConductorPipeline
|
28 |
+
from utils.gradio_utils import ensure_dirname, image2pil, split_filename, visualize_drag
|
29 |
from utils.lora_utils import add_LoRA_to_controlnet
|
30 |
+
from utils.utils import (
|
31 |
+
bivariate_Gaussian,
|
32 |
+
create_flow_controlnet,
|
33 |
+
create_image_controlnet,
|
34 |
+
interpolate_trajectory,
|
35 |
+
load_model,
|
36 |
+
load_weights,
|
37 |
+
save_videos_grid,
|
38 |
+
)
|
39 |
from utils.visualizer import Visualizer, vis_flow_to_video
|
40 |
+
|
41 |
#### Description ####
|
42 |
title = r"""<h1 align="center">CustomNet: Object Customization with Variable-Viewpoints in Text-to-Image Diffusion Models</h1>"""
|
43 |
|
|
|
49 |
<a href='https://liyaowei-stu.github.io/project/ImageConductor/'><img src='https://img.shields.io/badge/Project_Page-ImgaeConductor-green' alt='Project Page'></a>
|
50 |
<a href='https://arxiv.org/pdf/2406.15339'><img src='https://img.shields.io/badge/Paper-Arxiv-blue'></a>
|
51 |
<a href='https://github.com/liyaowei-stu/ImageConductor'><img src='https://img.shields.io/badge/Code-Github-orange'></a>
|
52 |
+
|
53 |
|
54 |
</div>
|
55 |
</br>
|
|
|
57 |
"""
|
58 |
|
59 |
|
|
|
60 |
descriptions = r"""
|
61 |
Official Gradio Demo for <a href='https://github.com/liyaowei-stu/ImageConductor'><b>Image Conductor: Precision Control for Interactive Video Synthesis</b></a>.<br>
|
62 |
🧙Image Conductor enables precise, fine-grained control for generating motion-controllable videos from images, advancing the practical application of interactive video synthesis.<br>
|
|
|
73 |
"""
|
74 |
|
75 |
citation = r"""
|
76 |
+
If Image Conductor is helpful, please help to ⭐ the <a href='https://github.com/liyaowei-stu/ImageConductor' target='_blank'>Github Repo</a>. Thanks!
|
77 |
[![GitHub Stars](https://img.shields.io/github/stars/liyaowei-stu%2FImageConductor)](https://github.com/liyaowei-stu/ImageConductor)
|
78 |
---
|
79 |
|
|
|
82 |
If our work is useful for your research, please consider citing:
|
83 |
```bibtex
|
84 |
@misc{li2024imageconductor,
|
85 |
+
title={Image Conductor: Precision Control for Interactive Video Synthesis},
|
86 |
author={Li, Yaowei and Wang, Xintao and Zhang, Zhaoyang and Wang, Zhouxia and Yuan, Ziyang and Xie, Liangbin and Zou, Yuexian and Shan, Ying},
|
87 |
year={2024},
|
88 |
eprint={2406.15339},
|
|
|
101 |
os.makedirs("models/sd1-5")
|
102 |
|
103 |
if not os.path.exists("models/flow_controlnet.ckpt"):
|
104 |
+
os.system(
|
105 |
+
f"wget -q https://huggingface.co/TencentARC/ImageConductor/resolve/main/flow_controlnet.ckpt?download=true -P models/"
|
106 |
+
)
|
107 |
+
os.system(f"mv models/flow_controlnet.ckpt?download=true models/flow_controlnet.ckpt")
|
108 |
+
print(
|
109 |
+
"flow_controlnet Download!",
|
110 |
+
)
|
111 |
|
112 |
if not os.path.exists("models/image_controlnet.ckpt"):
|
113 |
+
os.system(
|
114 |
+
f"wget -q https://huggingface.co/TencentARC/ImageConductor/resolve/main/image_controlnet.ckpt?download=true -P models/"
|
115 |
+
)
|
116 |
+
os.system(f"mv models/image_controlnet.ckpt?download=true models/image_controlnet.ckpt")
|
117 |
+
print(
|
118 |
+
"image_controlnet Download!",
|
119 |
+
)
|
120 |
|
121 |
if not os.path.exists("models/unet.ckpt"):
|
122 |
+
os.system(
|
123 |
+
f"wget -q https://huggingface.co/TencentARC/ImageConductor/resolve/main/unet.ckpt?download=true -P models/"
|
124 |
+
)
|
125 |
+
os.system(f"mv models/unet.ckpt?download=true models/unet.ckpt")
|
126 |
+
print(
|
127 |
+
"unet Download!",
|
128 |
+
)
|
129 |
+
|
130 |
|
|
|
131 |
if not os.path.exists("models/sd1-5/config.json"):
|
132 |
+
os.system(
|
133 |
+
f"wget -q https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/unet/config.json?download=true -P models/sd1-5/"
|
134 |
+
)
|
135 |
+
os.system(f"mv models/sd1-5/config.json?download=true models/sd1-5/config.json")
|
136 |
+
print(
|
137 |
+
"config Download!",
|
138 |
+
)
|
139 |
|
140 |
|
141 |
if not os.path.exists("models/sd1-5/unet.ckpt"):
|
142 |
+
os.system(f"cp -r models/unet.ckpt models/sd1-5/unet.ckpt")
|
143 |
|
144 |
# os.system(f'wget https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/unet/diffusion_pytorch_model.bin?download=true -P models/sd1-5/')
|
145 |
|
146 |
if not os.path.exists("models/personalized/helloobjects_V12c.safetensors"):
|
147 |
+
os.system(
|
148 |
+
f"wget -q https://huggingface.co/TencentARC/ImageConductor/resolve/main/helloobjects_V12c.safetensors?download=true -P models/personalized"
|
149 |
+
)
|
150 |
+
os.system(
|
151 |
+
f"mv models/personalized/helloobjects_V12c.safetensors?download=true models/personalized/helloobjects_V12c.safetensors"
|
152 |
+
)
|
153 |
+
print(
|
154 |
+
"helloobjects_V12c Download!",
|
155 |
+
)
|
156 |
|
157 |
|
158 |
if not os.path.exists("models/personalized/TUSUN.safetensors"):
|
159 |
+
os.system(
|
160 |
+
f"wget -q https://huggingface.co/TencentARC/ImageConductor/resolve/main/TUSUN.safetensors?download=true -P models/personalized"
|
161 |
+
)
|
162 |
+
os.system(f"mv models/personalized/TUSUN.safetensors?download=true models/personalized/TUSUN.safetensors")
|
163 |
+
print(
|
164 |
+
"TUSUN Download!",
|
165 |
+
)
|
166 |
|
167 |
# mv1 = os.system(f'mv /usr/local/lib/python3.10/site-packages/gradio/helpers.py /usr/local/lib/python3.10/site-packages/gradio/helpers_bkp.py')
|
168 |
# mv2 = os.system(f'mv helpers.py /usr/local/lib/python3.10/site-packages/gradio/helpers.py')
|
|
|
178 |
# - - - - - examples - - - - - #
|
179 |
|
180 |
image_examples = [
|
181 |
+
[
|
182 |
+
"__asset__/images/object/turtle-1.jpg",
|
183 |
+
"a sea turtle gracefully swimming over a coral reef in the clear blue ocean.",
|
184 |
+
"object",
|
185 |
+
11318446767408804497,
|
186 |
+
"",
|
187 |
+
"turtle",
|
188 |
+
"__asset__/turtle.mp4",
|
189 |
+
],
|
190 |
+
[
|
191 |
+
"__asset__/images/object/rose-1.jpg",
|
192 |
+
"a red rose engulfed in flames.",
|
193 |
+
"object",
|
194 |
+
6854275249656120509,
|
195 |
+
"",
|
196 |
+
"rose",
|
197 |
+
"__asset__/rose.mp4",
|
198 |
+
],
|
199 |
+
[
|
200 |
+
"__asset__/images/object/jellyfish-1.jpg",
|
201 |
+
"intricate detailing,photorealism,hyperrealistic, glowing jellyfish mushroom, flying, starry sky, bokeh, golden ratio composition.",
|
202 |
+
"object",
|
203 |
+
17966188172968903484,
|
204 |
+
"HelloObject",
|
205 |
+
"jellyfish",
|
206 |
+
"__asset__/jellyfish.mp4",
|
207 |
+
],
|
208 |
+
[
|
209 |
+
"__asset__/images/camera/lush-1.jpg",
|
210 |
+
"detailed craftsmanship, photorealism, hyperrealistic, roaring waterfall, misty spray, lush greenery, vibrant rainbow, golden ratio composition.",
|
211 |
+
"camera",
|
212 |
+
7970487946960948963,
|
213 |
+
"HelloObject",
|
214 |
+
"lush",
|
215 |
+
"__asset__/lush.mp4",
|
216 |
+
],
|
217 |
+
[
|
218 |
+
"__asset__/images/camera/tusun-1.jpg",
|
219 |
+
"tusuncub with its mouth open, blurry, open mouth, fangs, photo background, looking at viewer, tongue, full body, solo, cute and lovely, Beautiful and realistic eye details, perfect anatomy, Nonsense, pure background, Centered-Shot, realistic photo, photograph, 4k, hyper detailed, DSLR, 24 Megapixels, 8mm Lens, Full Frame, film grain, Global Illumination, studio Lighting, Award Winning Photography, diffuse reflection, ray tracing.",
|
220 |
+
"camera",
|
221 |
+
996953226890228361,
|
222 |
+
"TUSUN",
|
223 |
+
"tusun",
|
224 |
+
"__asset__/tusun.mp4",
|
225 |
+
],
|
226 |
+
[
|
227 |
+
"__asset__/images/camera/painting-1.jpg",
|
228 |
+
"A oil painting.",
|
229 |
+
"camera",
|
230 |
+
16867854766769816385,
|
231 |
+
"",
|
232 |
+
"painting",
|
233 |
+
"__asset__/painting.mp4",
|
234 |
+
],
|
235 |
]
|
236 |
|
237 |
|
238 |
POINTS = {
|
239 |
+
"turtle": "__asset__/trajs/object/turtle-1.json",
|
240 |
+
"rose": "__asset__/trajs/object/rose-1.json",
|
241 |
+
"jellyfish": "__asset__/trajs/object/jellyfish-1.json",
|
242 |
+
"lush": "__asset__/trajs/camera/lush-1.json",
|
243 |
+
"tusun": "__asset__/trajs/camera/tusun-1.json",
|
244 |
+
"painting": "__asset__/trajs/camera/painting-1.json",
|
245 |
}
|
246 |
|
247 |
IMAGE_PATH = {
|
248 |
+
"turtle": "__asset__/images/object/turtle-1.jpg",
|
249 |
+
"rose": "__asset__/images/object/rose-1.jpg",
|
250 |
+
"jellyfish": "__asset__/images/object/jellyfish-1.jpg",
|
251 |
+
"lush": "__asset__/images/camera/lush-1.jpg",
|
252 |
+
"tusun": "__asset__/images/camera/tusun-1.jpg",
|
253 |
+
"painting": "__asset__/images/camera/painting-1.jpg",
|
254 |
}
|
255 |
|
256 |
|
|
|
257 |
DREAM_BOOTH = {
|
258 |
+
"HelloObject": "models/personalized/helloobjects_V12c.safetensors",
|
259 |
}
|
260 |
|
261 |
LORA = {
|
262 |
+
"TUSUN": "models/personalized/TUSUN.safetensors",
|
263 |
}
|
264 |
|
265 |
LORA_ALPHA = {
|
266 |
+
"TUSUN": 0.6,
|
267 |
}
|
268 |
|
269 |
NPROMPT = {
|
270 |
+
"HelloObject": "FastNegativeV2,(bad-artist:1),(worst quality, low quality:1.4),(bad_prompt_version2:0.8),bad-hands-5,lowres,bad anatomy,bad hands,((text)),(watermark),error,missing fingers,extra digit,fewer digits,cropped,worst quality,low quality,normal quality,((username)),blurry,(extra limbs),bad-artist-anime,badhandv4,EasyNegative,ng_deepnegative_v1_75t,verybadimagenegative_v1.3,BadDream,(three hands:1.6),(three legs:1.2),(more than two hands:1.4),(more than two legs,:1.2)"
|
271 |
}
|
272 |
|
273 |
output_dir = "outputs"
|
274 |
ensure_dirname(output_dir)
|
275 |
|
276 |
+
|
277 |
def points_to_flows(track_points, model_length, height, width):
|
278 |
input_drag = np.zeros((model_length - 1, height, width, 2))
|
279 |
for splited_track in track_points:
|
280 |
+
if len(splited_track) == 1: # stationary point
|
281 |
displacement_point = tuple([splited_track[0][0] + 1, splited_track[0][1] + 1])
|
282 |
splited_track = tuple([splited_track[0], displacement_point])
|
283 |
# interpolate the track
|
284 |
splited_track = interpolate_trajectory(splited_track, model_length)
|
285 |
splited_track = splited_track[:model_length]
|
286 |
if len(splited_track) < model_length:
|
287 |
+
splited_track = splited_track + [splited_track[-1]] * (model_length - len(splited_track))
|
288 |
for i in range(model_length - 1):
|
289 |
start_point = splited_track[i]
|
290 |
+
end_point = splited_track[i + 1]
|
291 |
input_drag[i][int(start_point[1])][int(start_point[0])][0] = end_point[0] - start_point[0]
|
292 |
input_drag[i][int(start_point[1])][int(start_point[0])][1] = end_point[1] - start_point[1]
|
293 |
return input_drag
|
294 |
|
295 |
+
|
296 |
class ImageConductor:
|
297 |
+
def __init__(
|
298 |
+
self, device, unet_path, image_controlnet_path, flow_controlnet_path, height, width, model_length, lora_rank=64
|
299 |
+
):
|
300 |
self.device = device
|
301 |
+
tokenizer = CLIPTokenizer.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="tokenizer")
|
302 |
+
text_encoder = CLIPTextModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="text_encoder").to(
|
303 |
+
device
|
304 |
+
)
|
305 |
+
vae = AutoencoderKL.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="vae").to(device)
|
306 |
inference_config = OmegaConf.load("configs/inference/inference.yaml")
|
307 |
+
unet = UNet3DConditionFlowModel.from_pretrained_2d(
|
308 |
+
"models/sd1-5/", unet_additional_kwargs=OmegaConf.to_container(inference_config.unet_additional_kwargs)
|
309 |
+
)
|
310 |
|
311 |
self.vae = vae
|
312 |
|
|
|
327 |
|
328 |
self.pipeline = ImageConductorPipeline(
|
329 |
unet=unet,
|
330 |
+
vae=vae,
|
331 |
+
tokenizer=tokenizer,
|
332 |
+
text_encoder=text_encoder,
|
333 |
scheduler=DDIMScheduler(**OmegaConf.to_container(inference_config.noise_scheduler_kwargs)),
|
334 |
image_controlnet=image_controlnet,
|
335 |
flow_controlnet=flow_controlnet,
|
336 |
).to(device)
|
337 |
|
|
|
338 |
self.height = height
|
339 |
self.width = width
|
340 |
# _, model_step, _ = split_filename(model_path)
|
|
|
346 |
self.blur_kernel = blur_kernel
|
347 |
|
348 |
@spaces.GPU(duration=180)
|
349 |
+
def run(
|
350 |
+
self,
|
351 |
+
first_frame_path,
|
352 |
+
tracking_points,
|
353 |
+
prompt,
|
354 |
+
drag_mode,
|
355 |
+
negative_prompt,
|
356 |
+
seed,
|
357 |
+
randomize_seed,
|
358 |
+
guidance_scale,
|
359 |
+
num_inference_steps,
|
360 |
+
personalized,
|
361 |
+
examples_type,
|
362 |
+
):
|
363 |
print("Run!")
|
364 |
if examples_type != "":
|
365 |
### for adapting high version gradio
|
|
|
369 |
tracking_points.value.extend(points)
|
370 |
print("example first_frame_path", first_frame_path)
|
371 |
print("example tracking_points", tracking_points.value)
|
372 |
+
|
373 |
+
original_width, original_height = 384, 256
|
374 |
if isinstance(tracking_points, list):
|
375 |
input_all_points = tracking_points
|
376 |
else:
|
377 |
input_all_points = tracking_points.value
|
378 |
+
|
379 |
print("input_all_points", input_all_points)
|
380 |
+
resized_all_points = [
|
381 |
+
tuple(
|
382 |
+
[
|
383 |
+
tuple([float(e1[0] * self.width / original_width), float(e1[1] * self.height / original_height)])
|
384 |
+
for e1 in e
|
385 |
+
]
|
386 |
+
)
|
387 |
+
for e in input_all_points
|
388 |
+
]
|
389 |
|
390 |
dir, base, ext = split_filename(first_frame_path)
|
391 |
+
id = base.split("_")[-1]
|
392 |
+
|
393 |
+
visualized_drag, _ = visualize_drag(
|
394 |
+
first_frame_path, resized_all_points, drag_mode, self.width, self.height, self.model_length
|
395 |
+
)
|
396 |
|
397 |
+
## image condition
|
398 |
+
image_transforms = transforms.Compose(
|
399 |
+
[
|
400 |
transforms.RandomResizedCrop(
|
401 |
+
(self.height, self.width), (1.0, 1.0), ratio=(self.width / self.height, self.width / self.height)
|
|
|
402 |
),
|
403 |
transforms.ToTensor(),
|
404 |
+
]
|
405 |
+
)
|
406 |
|
407 |
image_paths = [first_frame_path]
|
408 |
controlnet_images = [(image_transforms(Image.open(path).convert("RGB"))) for path in image_paths]
|
|
|
411 |
num_controlnet_images = controlnet_images.shape[2]
|
412 |
controlnet_images = rearrange(controlnet_images, "b c f h w -> (b f) c h w")
|
413 |
self.vae.to(device)
|
414 |
+
controlnet_images = self.vae.encode(controlnet_images * 2.0 - 1.0).latent_dist.sample() * 0.18215
|
415 |
controlnet_images = rearrange(controlnet_images, "(b f) c h w -> b c f h w", f=num_controlnet_images)
|
416 |
|
417 |
# flow condition
|
418 |
controlnet_flows = points_to_flows(resized_all_points, self.model_length, self.height, self.width)
|
419 |
+
for i in range(0, self.model_length - 1):
|
420 |
controlnet_flows[i] = cv2.filter2D(controlnet_flows[i], -1, self.blur_kernel)
|
421 |
+
controlnet_flows = np.concatenate(
|
422 |
+
[np.zeros_like(controlnet_flows[0])[np.newaxis, ...], controlnet_flows], axis=0
|
423 |
+
) # pad the first frame with zero flow
|
424 |
os.makedirs(os.path.join(output_dir, "control_flows"), exist_ok=True)
|
425 |
+
trajs_video = vis_flow_to_video(controlnet_flows, num_frames=self.model_length) # T-1 x H x W x 3
|
426 |
+
torchvision.io.write_video(
|
427 |
+
f"{output_dir}/control_flows/sample-{id}-train_flow.mp4",
|
428 |
+
trajs_video,
|
429 |
+
fps=8,
|
430 |
+
video_codec="h264",
|
431 |
+
options={"crf": "10"},
|
432 |
+
)
|
433 |
+
controlnet_flows = torch.from_numpy(controlnet_flows)[None][:, : self.model_length, ...]
|
434 |
+
controlnet_flows = rearrange(controlnet_flows, "b f h w c-> b c f h w").float().to(device)
|
435 |
|
436 |
+
dreambooth_model_path = DREAM_BOOTH.get(personalized, "")
|
437 |
+
lora_model_path = LORA.get(personalized, "")
|
438 |
lora_alpha = LORA_ALPHA.get(personalized, 0.6)
|
439 |
self.pipeline = load_weights(
|
440 |
self.pipeline,
|
441 |
+
dreambooth_model_path=dreambooth_model_path,
|
442 |
+
lora_model_path=lora_model_path,
|
443 |
+
lora_alpha=lora_alpha,
|
444 |
).to(device)
|
445 |
+
|
446 |
+
if NPROMPT.get(personalized, "") != "":
|
447 |
+
negative_prompt = NPROMPT.get(personalized)
|
448 |
+
|
449 |
if randomize_seed:
|
450 |
random_seed = torch.seed()
|
451 |
else:
|
452 |
seed = int(seed)
|
453 |
random_seed = seed
|
454 |
torch.manual_seed(random_seed)
|
455 |
+
torch.cuda.manual_seed_all(random_seed)
|
456 |
print(f"current seed: {torch.initial_seed()}")
|
457 |
sample = self.pipeline(
|
458 |
+
prompt,
|
459 |
+
negative_prompt=negative_prompt,
|
460 |
+
num_inference_steps=num_inference_steps,
|
461 |
+
guidance_scale=guidance_scale,
|
462 |
+
width=self.width,
|
463 |
+
height=self.height,
|
464 |
+
video_length=self.model_length,
|
465 |
+
controlnet_images=controlnet_images, # 1 4 1 32 48
|
466 |
+
controlnet_image_index=[0],
|
467 |
+
controlnet_flows=controlnet_flows, # [1, 2, 16, 256, 384]
|
468 |
+
control_mode=drag_mode,
|
469 |
+
eval_mode=True,
|
470 |
+
).videos
|
471 |
+
|
472 |
+
outputs_path = os.path.join(output_dir, f"output_{i}_{id}.mp4")
|
473 |
+
vis_video = (rearrange(sample[0], "c t h w -> t h w c") * 255.0).clip(0, 255)
|
474 |
+
torchvision.io.write_video(outputs_path, vis_video, fps=8, video_codec="h264", options={"crf": "10"})
|
475 |
+
|
476 |
# outputs_path = os.path.join(output_dir, f'output_{i}_{id}.gif')
|
477 |
# save_videos_grid(sample[0][None], outputs_path)
|
478 |
print("Done!")
|
|
|
482 |
def reset_states(first_frame_path, tracking_points):
|
483 |
first_frame_path = gr.State()
|
484 |
tracking_points = gr.State([])
|
485 |
+
return {input_image: None, first_frame_path_var: first_frame_path, tracking_points_var: tracking_points}
|
486 |
|
487 |
|
488 |
def preprocess_image(image, tracking_points):
|
489 |
image_pil = image2pil(image.name)
|
490 |
raw_w, raw_h = image_pil.size
|
491 |
+
resize_ratio = max(384 / raw_w, 256 / raw_h)
|
492 |
image_pil = image_pil.resize((int(raw_w * resize_ratio), int(raw_h * resize_ratio)), Image.BILINEAR)
|
493 |
+
image_pil = transforms.CenterCrop((256, 384))(image_pil.convert("RGB"))
|
494 |
id = str(uuid.uuid4())[:4]
|
495 |
first_frame_path = os.path.join(output_dir, f"first_frame_{id}.jpg")
|
496 |
image_pil.save(first_frame_path, quality=95)
|
497 |
+
tracking_points = gr.State([])
|
498 |
+
return {
|
499 |
+
input_image: first_frame_path,
|
500 |
+
first_frame_path_var: first_frame_path,
|
501 |
+
tracking_points_var: tracking_points,
|
502 |
+
personalized: "",
|
503 |
+
}
|
504 |
+
|
505 |
+
|
506 |
+
def add_tracking_points(
|
507 |
+
tracking_points, first_frame_path, drag_mode, evt: gr.SelectData
|
508 |
+
): # SelectData is a subclass of EventData
|
509 |
+
if drag_mode == "object":
|
510 |
color = (255, 0, 0, 255)
|
511 |
+
elif drag_mode == "camera":
|
512 |
color = (0, 0, 255, 255)
|
513 |
|
514 |
+
if not isinstance(tracking_points, list):
|
515 |
print(f"You selected {evt.value} at {evt.index} from {evt.target}")
|
516 |
tracking_points.value[-1].append(evt.index)
|
517 |
print(tracking_points.value)
|
518 |
+
tracking_points_values = tracking_points.value
|
519 |
else:
|
520 |
try:
|
521 |
tracking_points[-1].append(evt.index)
|
|
|
523 |
tracking_points.append([])
|
524 |
tracking_points[-1].append(evt.index)
|
525 |
print(f"Solved Error: {e}")
|
526 |
+
|
527 |
tracking_points_values = tracking_points
|
528 |
+
|
529 |
+
transparent_background = Image.open(first_frame_path).convert("RGBA")
|
|
|
530 |
w, h = transparent_background.size
|
531 |
transparent_layer = np.zeros((h, w, 4))
|
532 |
+
|
533 |
for track in tracking_points_values:
|
534 |
if len(track) > 1:
|
535 |
+
for i in range(len(track) - 1):
|
536 |
start_point = track[i]
|
537 |
+
end_point = track[i + 1]
|
538 |
vx = end_point[0] - start_point[0]
|
539 |
vy = end_point[1] - start_point[1]
|
540 |
arrow_length = np.sqrt(vx**2 + vy**2)
|
541 |
+
if i == len(track) - 2:
|
542 |
+
cv2.arrowedLine(
|
543 |
+
transparent_layer, tuple(start_point), tuple(end_point), color, 2, tipLength=8 / arrow_length
|
544 |
+
)
|
545 |
else:
|
546 |
+
cv2.line(
|
547 |
+
transparent_layer,
|
548 |
+
tuple(start_point),
|
549 |
+
tuple(end_point),
|
550 |
+
color,
|
551 |
+
2,
|
552 |
+
)
|
553 |
else:
|
554 |
cv2.circle(transparent_layer, tuple(track[0]), 5, color, -1)
|
555 |
|
|
|
559 |
|
560 |
|
561 |
def add_drag(tracking_points):
|
562 |
+
if not isinstance(tracking_points, list):
|
563 |
# print("before", tracking_points.value)
|
564 |
tracking_points.value.append([])
|
565 |
# print(tracking_points.value)
|
566 |
else:
|
567 |
tracking_points.append([])
|
568 |
return {tracking_points_var: tracking_points}
|
569 |
+
|
570 |
|
571 |
def delete_last_drag(tracking_points, first_frame_path, drag_mode):
|
572 |
+
if drag_mode == "object":
|
573 |
color = (255, 0, 0, 255)
|
574 |
+
elif drag_mode == "camera":
|
575 |
color = (0, 0, 255, 255)
|
576 |
tracking_points.value.pop()
|
577 |
+
transparent_background = Image.open(first_frame_path).convert("RGBA")
|
578 |
w, h = transparent_background.size
|
579 |
transparent_layer = np.zeros((h, w, 4))
|
580 |
for track in tracking_points.value:
|
581 |
if len(track) > 1:
|
582 |
+
for i in range(len(track) - 1):
|
583 |
start_point = track[i]
|
584 |
+
end_point = track[i + 1]
|
585 |
vx = end_point[0] - start_point[0]
|
586 |
vy = end_point[1] - start_point[1]
|
587 |
arrow_length = np.sqrt(vx**2 + vy**2)
|
588 |
+
if i == len(track) - 2:
|
589 |
+
cv2.arrowedLine(
|
590 |
+
transparent_layer, tuple(start_point), tuple(end_point), color, 2, tipLength=8 / arrow_length
|
591 |
+
)
|
592 |
else:
|
593 |
+
cv2.line(
|
594 |
+
transparent_layer,
|
595 |
+
tuple(start_point),
|
596 |
+
tuple(end_point),
|
597 |
+
color,
|
598 |
+
2,
|
599 |
+
)
|
600 |
else:
|
601 |
cv2.circle(transparent_layer, tuple(track[0]), 5, color, -1)
|
602 |
|
603 |
transparent_layer = Image.fromarray(transparent_layer.astype(np.uint8))
|
604 |
trajectory_map = Image.alpha_composite(transparent_background, transparent_layer)
|
605 |
return {tracking_points_var: tracking_points, input_image: trajectory_map}
|
606 |
+
|
607 |
|
608 |
def delete_last_step(tracking_points, first_frame_path, drag_mode):
|
609 |
+
if drag_mode == "object":
|
610 |
color = (255, 0, 0, 255)
|
611 |
+
elif drag_mode == "camera":
|
612 |
color = (0, 0, 255, 255)
|
613 |
tracking_points.value[-1].pop()
|
614 |
+
transparent_background = Image.open(first_frame_path).convert("RGBA")
|
615 |
w, h = transparent_background.size
|
616 |
transparent_layer = np.zeros((h, w, 4))
|
617 |
for track in tracking_points.value:
|
618 |
if len(track) > 1:
|
619 |
+
for i in range(len(track) - 1):
|
620 |
start_point = track[i]
|
621 |
+
end_point = track[i + 1]
|
622 |
vx = end_point[0] - start_point[0]
|
623 |
vy = end_point[1] - start_point[1]
|
624 |
arrow_length = np.sqrt(vx**2 + vy**2)
|
625 |
+
if i == len(track) - 2:
|
626 |
+
cv2.arrowedLine(
|
627 |
+
transparent_layer, tuple(start_point), tuple(end_point), color, 2, tipLength=8 / arrow_length
|
628 |
+
)
|
629 |
else:
|
630 |
+
cv2.line(
|
631 |
+
transparent_layer,
|
632 |
+
tuple(start_point),
|
633 |
+
tuple(end_point),
|
634 |
+
color,
|
635 |
+
2,
|
636 |
+
)
|
637 |
else:
|
638 |
+
cv2.circle(transparent_layer, tuple(track[0]), 5, color, -1)
|
639 |
|
640 |
transparent_layer = Image.fromarray(transparent_layer.astype(np.uint8))
|
641 |
trajectory_map = Image.alpha_composite(transparent_background, transparent_layer)
|
642 |
return {tracking_points_var: tracking_points, input_image: trajectory_map}
|
643 |
|
644 |
|
645 |
+
block = gr.Blocks(theme=gr.themes.Soft(radius_size=gr.themes.sizes.radius_none, text_size=gr.themes.sizes.text_md))
|
|
|
|
|
|
|
|
|
|
|
646 |
with block:
|
647 |
with gr.Row():
|
648 |
with gr.Column():
|
|
|
652 |
|
653 |
with gr.Accordion(label="🛠️ Instructions:", open=True, elem_id="accordion"):
|
654 |
with gr.Row(equal_height=True):
|
655 |
+
gr.Markdown(instructions)
|
|
|
656 |
|
657 |
# device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
658 |
device = torch.device("cuda")
|
659 |
+
unet_path = "models/unet.ckpt"
|
660 |
+
image_controlnet_path = "models/image_controlnet.ckpt"
|
661 |
+
flow_controlnet_path = "models/flow_controlnet.ckpt"
|
662 |
+
ImageConductor_net = ImageConductor(
|
663 |
+
device=device,
|
664 |
+
unet_path=unet_path,
|
665 |
+
image_controlnet_path=image_controlnet_path,
|
666 |
+
flow_controlnet_path=flow_controlnet_path,
|
667 |
+
height=256,
|
668 |
+
width=384,
|
669 |
+
model_length=16,
|
670 |
+
)
|
671 |
first_frame_path_var = gr.State(value=None)
|
672 |
tracking_points_var = gr.State([])
|
673 |
|
674 |
with gr.Row():
|
675 |
with gr.Column(scale=1):
|
676 |
+
image_upload_button = gr.UploadButton(label="Upload Image", file_types=["image"])
|
677 |
add_drag_button = gr.Button(value="Add Drag")
|
678 |
reset_button = gr.Button(value="Reset")
|
679 |
delete_last_drag_button = gr.Button(value="Delete last drag")
|
680 |
delete_last_step_button = gr.Button(value="Delete last step")
|
|
|
|
|
681 |
|
682 |
with gr.Column(scale=7):
|
683 |
with gr.Row():
|
684 |
with gr.Column(scale=6):
|
685 |
+
input_image = gr.Image(
|
686 |
+
label="Input Image",
|
687 |
+
interactive=True,
|
688 |
+
height=300,
|
689 |
+
width=384,
|
690 |
+
)
|
691 |
with gr.Column(scale=6):
|
692 |
+
output_image = gr.Image(
|
693 |
+
label="Motion Path",
|
694 |
+
interactive=False,
|
695 |
+
height=256,
|
696 |
+
width=384,
|
697 |
+
)
|
698 |
with gr.Row():
|
699 |
with gr.Column(scale=1):
|
700 |
+
prompt = gr.Textbox(
|
701 |
+
value="a wonderful elf.", label="Prompt (highly-recommended)", interactive=True, visible=True
|
702 |
+
)
|
703 |
negative_prompt = gr.Text(
|
704 |
+
label="Negative Prompt",
|
705 |
+
max_lines=5,
|
706 |
+
placeholder="Please input your negative prompt",
|
707 |
+
value="worst quality, low quality, letterboxed",
|
708 |
+
lines=1,
|
709 |
+
)
|
710 |
+
drag_mode = gr.Radio(["camera", "object"], label="Drag mode: ", value="object", scale=2)
|
711 |
run_button = gr.Button(value="Run")
|
712 |
|
713 |
with gr.Accordion("More input params", open=False, elem_id="accordion1"):
|
714 |
with gr.Group():
|
715 |
seed = gr.Textbox(
|
716 |
+
label="Seed: ",
|
717 |
+
value=561793204,
|
718 |
)
|
719 |
randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
|
720 |
+
|
721 |
with gr.Group():
|
722 |
with gr.Row():
|
723 |
guidance_scale = gr.Slider(
|
|
|
734 |
step=1,
|
735 |
value=25,
|
736 |
)
|
737 |
+
|
738 |
with gr.Group():
|
739 |
+
personalized = gr.Dropdown(label="Personalized", choices=["", "HelloObject", "TUSUN"], value="")
|
740 |
+
examples_type = gr.Textbox(label="Examples Type (Ignore) ", value="", visible=False)
|
741 |
|
742 |
with gr.Column(scale=7):
|
743 |
+
output_video = gr.Video(label="Output Video", width=384, height=256)
|
|
|
|
|
|
|
744 |
# output_video = gr.Image(label="Output Video",
|
745 |
# height=256,
|
746 |
# width=384,)
|
747 |
+
|
|
|
748 |
with gr.Row():
|
|
|
749 |
|
750 |
example = gr.Examples(
|
751 |
label="Input Example",
|
|
|
754 |
examples_per_page=10,
|
755 |
cache_examples=False,
|
756 |
)
|
757 |
+
|
|
|
758 |
with gr.Row():
|
759 |
gr.Markdown(citation)
|
760 |
|
761 |
+
image_upload_button.upload(
|
762 |
+
preprocess_image,
|
763 |
+
[image_upload_button, tracking_points_var],
|
764 |
+
[input_image, first_frame_path_var, tracking_points_var, personalized],
|
765 |
+
)
|
766 |
|
767 |
add_drag_button.click(add_drag, tracking_points_var, tracking_points_var)
|
768 |
|
769 |
+
delete_last_drag_button.click(
|
770 |
+
delete_last_drag, [tracking_points_var, first_frame_path_var, drag_mode], [tracking_points_var, input_image]
|
771 |
+
)
|
772 |
+
|
773 |
+
delete_last_step_button.click(
|
774 |
+
delete_last_step, [tracking_points_var, first_frame_path_var, drag_mode], [tracking_points_var, input_image]
|
775 |
+
)
|
776 |
+
|
777 |
+
reset_button.click(
|
778 |
+
reset_states,
|
779 |
+
[first_frame_path_var, tracking_points_var],
|
780 |
+
[input_image, first_frame_path_var, tracking_points_var],
|
781 |
+
)
|
782 |
+
|
783 |
+
input_image.select(
|
784 |
+
add_tracking_points, [tracking_points_var, first_frame_path_var, drag_mode], [tracking_points_var, input_image]
|
785 |
+
)
|
786 |
+
|
787 |
+
run_button.click(
|
788 |
+
ImageConductor_net.run,
|
789 |
+
[
|
790 |
+
first_frame_path_var,
|
791 |
+
tracking_points_var,
|
792 |
+
prompt,
|
793 |
+
drag_mode,
|
794 |
+
negative_prompt,
|
795 |
+
seed,
|
796 |
+
randomize_seed,
|
797 |
+
guidance_scale,
|
798 |
+
num_inference_steps,
|
799 |
+
personalized,
|
800 |
+
examples_type,
|
801 |
+
],
|
802 |
+
[output_image, output_video],
|
803 |
+
)
|
804 |
|
805 |
block.queue().launch()
|
requirements.txt
CHANGED
@@ -1,29 +1,28 @@
|
|
1 |
-
torch
|
2 |
-
torchvision
|
3 |
-
torchaudio
|
4 |
-
transformers==4.32.1
|
5 |
-
gradio==4.38.1
|
6 |
-
ftfy
|
7 |
-
tensorboard
|
8 |
-
datasets
|
9 |
-
Pillow==9.5.0
|
10 |
-
opencv-python
|
11 |
-
imgaug
|
12 |
accelerate==0.23.0
|
13 |
-
image-reward
|
14 |
-
hpsv2
|
15 |
-
torchmetrics
|
16 |
-
open-clip-torch
|
17 |
-
clip
|
18 |
av2
|
19 |
-
|
20 |
-
|
21 |
-
scipy
|
22 |
-
tqdm
|
23 |
-
einops
|
24 |
diffusers==0.28.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
omegaconf
|
|
|
|
|
|
|
|
|
26 |
scikit-image
|
27 |
scikit-learn
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
accelerate==0.23.0
|
|
|
|
|
|
|
|
|
|
|
2 |
av2
|
3 |
+
clip
|
4 |
+
datasets
|
|
|
|
|
|
|
5 |
diffusers==0.28.0
|
6 |
+
einops
|
7 |
+
ftfy
|
8 |
+
gradio==4.38.1
|
9 |
+
hpsv2
|
10 |
+
image-reward
|
11 |
+
imageio-ffmpeg
|
12 |
+
imgaug
|
13 |
+
numpy==1.26.2
|
14 |
omegaconf
|
15 |
+
open-clip-torch
|
16 |
+
opencv-python
|
17 |
+
peft
|
18 |
+
Pillow==9.5.0
|
19 |
scikit-image
|
20 |
scikit-learn
|
21 |
+
scipy
|
22 |
+
tensorboard
|
23 |
+
torch
|
24 |
+
torchaudio
|
25 |
+
torchmetrics
|
26 |
+
torchvision
|
27 |
+
tqdm
|
28 |
+
transformers==4.32.1
|