File size: 11,829 Bytes
e79d24a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3d3780a
eef7e90
 
 
 
 
e79d24a
 
 
 
 
 
 
 
 
 
eef7e90
e79d24a
eef7e90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0623638
eef7e90
 
 
 
 
 
 
 
 
 
 
 
 
 
3d3780a
eef7e90
 
 
 
 
 
 
 
 
 
e79d24a
eef7e90
 
 
 
 
 
 
 
 
 
 
 
 
e79d24a
eef7e90
e79d24a
eef7e90
e79d24a
 
 
 
 
 
 
 
 
 
 
eef7e90
e79d24a
 
 
 
 
eef7e90
 
 
 
e79d24a
eef7e90
e79d24a
 
 
 
 
 
 
 
 
 
4157194
 
e79d24a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
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
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
import os
import cv2
import torch
import gradio as gr
import torchvision
import warnings
import numpy as np
from PIL import Image, ImageSequence
from moviepy.editor import VideoFileClip
import imageio
from diffusers import (
    TextToVideoSDPipeline,
    AutoencoderKL,
    DDPMScheduler,
    DDIMScheduler,
    UNet3DConditionModel,
)
from transformers import CLIPTokenizer, CLIPTextModel
from diffusers.utils import export_to_video
from typing import List
from text2vid_modded import TextToVideoSDPipelineModded
from invert_utils import ddim_inversion as dd_inversion
from gifs_filter import filter
import subprocess
import spaces


def load_frames(image: Image, mode='RGBA'):
    return np.array([np.array(frame.convert(mode)) for frame in ImageSequence.Iterator(image)])


def run_setup():
    try:
        # Step 1: Install Git LFS
        subprocess.run(["git", "lfs", "install"], check=True)

        # Step 2: Clone the repository
        repo_url = "https://huggingface.co/Hmrishav/t2v_sketch-lora"
        subprocess.run(["git", "clone", repo_url], check=True)

        # Step 3: Move the checkpoint file
        source = "t2v_sketch-lora/checkpoint-2500"
        destination = "./checkpoint-2500/"
        os.rename(source, destination)

        print("Setup completed successfully!")
    except subprocess.CalledProcessError as e:
        print(f"Error during setup: {e}")
    except FileNotFoundError as e:
        print(f"File operation error: {e}")
    except Exception as e:
        print(f"Unexpected error: {e}")

# Automatically run setup during app initialization
run_setup()


def save_gif(frames, path):
    imageio.mimsave(
        path,
        [frame.astype(np.uint8) for frame in frames],
        format="GIF",
        duration=1 / 10,
        loop=0  # 0 means infinite loop
    )

def load_image(imgname, target_size=None):
    pil_img = Image.open(imgname).convert('RGB')
    if target_size:
        if isinstance(target_size, int):
            target_size = (target_size, target_size)
        pil_img = pil_img.resize(target_size, Image.Resampling.LANCZOS)
    return torchvision.transforms.ToTensor()(pil_img).unsqueeze(0)

def prepare_latents(pipe, x_aug):
    with torch.cuda.amp.autocast():
        batch_size, num_frames, channels, height, width = x_aug.shape
        x_aug = x_aug.reshape(batch_size * num_frames, channels, height, width)
        latents = pipe.vae.encode(x_aug).latent_dist.sample()
        latents = latents.view(batch_size, num_frames, -1, latents.shape[2], latents.shape[3])
        latents = latents.permute(0, 2, 1, 3, 4)
    return pipe.vae.config.scaling_factor * latents


@torch.no_grad()
def invert(pipe, inv, load_name, device="cuda", dtype=torch.bfloat16):
    input_img = [load_image(load_name, 256).to(device, dtype=dtype).unsqueeze(1)] * 5
    input_img = torch.cat(input_img, dim=1)
    latents = prepare_latents(pipe, input_img).to(torch.bfloat16)
    inv.set_timesteps(25)
    id_latents = dd_inversion(pipe, inv, video_latent=latents, num_inv_steps=25, prompt="")[-1].to(dtype)
    return torch.mean(id_latents, dim=2, keepdim=True)

def load_primary_models(pretrained_model_path):
    return (
        DDPMScheduler.from_config(pretrained_model_path, subfolder="scheduler"),
        CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer"),
        CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder"),
        AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae"),
        UNet3DConditionModel.from_pretrained(pretrained_model_path, subfolder="unet"),
    )

def initialize_pipeline(model: str, device: str = "cuda"):
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        scheduler, tokenizer, text_encoder, vae, unet = load_primary_models(model)
    pipe = TextToVideoSDPipeline.from_pretrained(
        pretrained_model_name_or_path="damo-vilab/text-to-video-ms-1.7b",
        scheduler=scheduler,
        tokenizer=tokenizer,
        text_encoder=text_encoder.to(device=device, dtype=torch.bfloat16),
        vae=vae.to(device=device, dtype=torch.bfloat16),
        unet=unet.to(device=device, dtype=torch.bfloat16),
    )
    pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
    return pipe, pipe.scheduler

# Initialize the models
LORA_CHECKPOINT = "checkpoint-2500"
os.environ["TORCH_CUDNN_V8_API_ENABLED"] = "1"
device = 'cuda' if torch.cuda.is_available() else 'cpu'
dtype = torch.bfloat16

pipe_inversion, inv = initialize_pipeline(LORA_CHECKPOINT, device)
pipe = TextToVideoSDPipelineModded.from_pretrained(
    pretrained_model_name_or_path="damo-vilab/text-to-video-ms-1.7b",
    scheduler=pipe_inversion.scheduler,
    tokenizer=pipe_inversion.tokenizer,
    text_encoder=pipe_inversion.text_encoder,
    vae=pipe_inversion.vae,
    unet=pipe_inversion.unet,
).to(device)

@spaces.GPU(duration=100)
@torch.no_grad()
def process_video(num_frames, num_seeds, generator, exp_dir, load_name, caption, lambda_):
    pipe_inversion.to(device)
    id_latents = invert(pipe_inversion, inv, load_name).to(device, dtype=dtype)
    latents = id_latents.repeat(num_seeds, 1, 1, 1, 1)
    generator = [torch.Generator(device="cuda").manual_seed(i) for i in range(num_seeds)]
    video_frames = pipe(
        prompt=caption,
        negative_prompt="",
        num_frames=num_frames,
        num_inference_steps=25,
        inv_latents=latents,
        guidance_scale=9,
        generator=generator,
        lambda_=lambda_,
    ).frames

    gifs = []
    for seed in range(num_seeds):
        vid_name = f"{exp_dir}/mp4_logs/vid_{os.path.basename(load_name)[:-4]}-rand{seed}.mp4"
        gif_name = f"{exp_dir}/gif_logs/vid_{os.path.basename(load_name)[:-4]}-rand{seed}.gif"
        
        os.makedirs(os.path.dirname(vid_name), exist_ok=True)
        os.makedirs(os.path.dirname(gif_name), exist_ok=True)
        
        video_path = export_to_video(video_frames[seed], output_video_path=vid_name)
        VideoFileClip(vid_name).write_gif(gif_name)
        
        with Image.open(gif_name) as im:
            frames = load_frames(im)

        frames_collect = np.empty((0, 1024, 1024), int)
        for frame in frames:
            frame = cv2.resize(frame, (1024, 1024))[:, :, :3]
            frame = cv2.cvtColor(255 - frame, cv2.COLOR_RGB2GRAY)
            _, frame = cv2.threshold(255 - frame, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
            frames_collect = np.append(frames_collect, [frame], axis=0)

        save_gif(frames_collect, gif_name)
        gifs.append(gif_name)

    return gifs

def generate_output(image, prompt: str, num_seeds: int = 3, lambda_value: float = 0.5) -> List[str]:
    """Main function to generate output GIFs"""
    exp_dir = "static/app_tmp"
    os.makedirs(exp_dir, exist_ok=True)
    
    # Save the input image temporarily
    temp_image_path = os.path.join(exp_dir, "temp_input.png")
    image.save(temp_image_path)
    
    # Generate the GIFs
    generated_gifs = process_video(
        num_frames=10,
        num_seeds=num_seeds,
        generator=None,
        exp_dir=exp_dir,
        load_name=temp_image_path,
        caption=prompt,
        lambda_=1 - lambda_value
    )
    
    # Apply filtering (assuming filter function is imported)
    filtered_gifs = filter(generated_gifs, temp_image_path)
    
    return filtered_gifs

css=""" """

def create_gradio_interface():
    with gr.Blocks(css=css) as demo:
        with gr.Column():
            gr.Markdown(
            """
                        
            <div align="center" id = "user-content-toc">
            <img align="left" width="70" height="70" src="https://github.com/user-attachments/assets/c61cec76-3c4b-42eb-8c65-f07e0166b7d8" alt="">
            
            # [FlipSketch: Flipping assets Drawings to Text-Guided Sketch Animations](https://hmrishavbandy.github.io/flipsketch-web/)
            ## [Hmrishav Bandyopadhyay](https://hmrishavbandy.github.io/) . [Yi-Zhe Song](https://personalpages.surrey.ac.uk/y.song/)
            </div>

            """
            )
        
        with gr.Row():
            with gr.Column():        
                
                input_sketch = gr.Image(
                    type="pil",
                    label="Selected Sketch",
                    scale=1,
                    interactive=True,
                    height=300  # Fixed height for consistency
                )
               
                motion_prompt = gr.Textbox(
                    label="Prompt",
                    placeholder="Describe the motion...",
                    lines=3
                )

                gr.Examples(
                    examples=[
                        ['./static/examples/sketch1.png', 'The camel walks slowly'],
                        ['./static/examples/sketch2.png', 'The wine in the wine glass sways from side to side'],
                        ['./static/examples/sketch3.png', 'The squirrel is eating a nut'],
                        ['./static/examples/sketch4.png', 'The surfer surfs on the waves'],
                        ['./static/examples/sketch5.png', 'A galloping horse'],
                        ['./static/examples/sketch6.png', 'The cat walks forward'],
                        ['./static/examples/sketch7.png', 'The eagle flies in the sky'],
                        ['./static/examples/sketch8.png', 'The flower is blooming slowly'],
                        ['./static/examples/sketch9.png', 'The reindeer looks around'],
                        ['./static/examples/sketch10.png', 'The cloud floats in the sky'],
                        ['./static/examples/sketch11.png', 'The jazz saxophonist performs on stage with a rhythmic sway, his upper body sways subtly to the rhythm of the music.'],
                        ['./static/examples/sketch12.png', 'The biker rides on the road']
                   ],
                   inputs=[input_sketch, motion_prompt],
                   examples_per_page=1
               )
                
                with gr.Row():
                    num_seeds = gr.Slider(
                        minimum=1, 
                        maximum=10, 
                        value=5, 
                        step=1, 
                        label="Seeds"
                    )
                    lambda_ = gr.Slider(
                        minimum=0, 
                        maximum=1, 
                        value=0.5, 
                        step=0.1, 
                        label="Motion Strength"
                    )
                generate_btn = gr.Button(
                    "Generate Animation",
                    variant="primary",
                    elem_classes="generate-btn",
                    interactive=True,
                )

            with gr.Column():              
                
                output_gallery = gr.Gallery(
                    label="Results",
                    elem_classes="output-gallery",
                    columns=3,
                    rows=2,
                    height="auto",
                    allow_preview=False,  # Disable preview expansion
                    show_share_button=False,
                    object_fit="cover",
                    preview=False
                )
            
           
        
        # Event handlers
        example_generate_btn.click(
            fn=generate_output,
            inputs=[
                input_sketch,
                motion_prompt,
                num_seeds,
                lambda_
            ],
            outputs=output_gallery
        )
    
    return demo

# Launch the app
if __name__ == "__main__":
    demo = create_gradio_interface()
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        show_api=False,
        ssr_mode=False
    )