xi0v Fabrice-TIERCELIN commited on
Commit
89df602
1 Parent(s): 4f5581d

Reuse the code of a working space (Python code) (#6)

Browse files

- Reuse the code of a working space (Python code) (0dbf91c605718ef47c74ecae87adce105e776a50)


Co-authored-by: Fabrice TIERCELIN <Fabrice-TIERCELIN@users.noreply.huggingface.co>

Files changed (1) hide show
  1. app.py +96 -31
app.py CHANGED
@@ -1,44 +1,109 @@
1
  import gradio as gr
2
- from diffusers import DiffusionPipeline
3
- from diffusers.utils import export_to_video
4
  import torch
5
  import os
 
 
 
 
 
 
6
  from PIL import Image
 
 
 
 
7
  import spaces
 
8
 
9
- # Load the pre-trained pipeline
10
- pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt")
11
-
12
- # Define the Gradio interface
13
- interface = gr.Interface(
14
- fn=lambda img: generate_video(img),
15
- inputs=gr.Image(type="pil"),
16
- outputs=gr.Video(),
17
- title="Stable Video Diffusion",
18
- description="Upload an image to generate a video",
19
- theme="soft"
20
  )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
 
22
- @spaces.GPU(duration=360)
23
- def generate_video(image):
24
- """
25
- Generates a video from an input image using the pipeline.
 
26
 
27
- Args:
28
- image: A PIL Image object representing the input image.
 
 
29
 
30
- Returns:
31
- The path of a video file.
32
- """
33
- video_frames = pipeline(image=image, num_inference_steps=10).images
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
 
35
- # Frames to Video
36
- os.makedirs("outputs", exist_ok=True)
37
- base_count = len(glob(os.path.join("outputs", "*.mp4")))
38
- video_path = os.path.join("outputs", f"{base_count:06d}.mp4")
39
- export_to_video(video_frames, video_path, fps=6)
40
 
41
- return video_path
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42
 
43
- # Launch the Gradio app
44
- interface.launch()
 
 
1
  import gradio as gr
2
+ #import gradio.helpers
 
3
  import torch
4
  import os
5
+ from glob import glob
6
+ from pathlib import Path
7
+ from typing import Optional
8
+
9
+ from diffusers import StableVideoDiffusionPipeline
10
+ from diffusers.utils import load_image, export_to_video
11
  from PIL import Image
12
+
13
+ import uuid
14
+ import random
15
+ from huggingface_hub import hf_hub_download
16
  import spaces
17
+ #gradio.helpers.CACHED_FOLDER = '/data/cache'
18
 
19
+ pipe = StableVideoDiffusionPipeline.from_pretrained(
20
+ "multimodalart/stable-video-diffusion", torch_dtype=torch.float16, variant="fp16"
 
 
 
 
 
 
 
 
 
21
  )
22
+ pipe.to("cuda")
23
+ #pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
24
+ #pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=True)
25
+
26
+ max_64_bit_int = 2**63 - 1
27
+
28
+ @spaces.GPU(duration=120)
29
+ def sample(
30
+ image: Image,
31
+ seed: Optional[int] = 42,
32
+ randomize_seed: bool = True,
33
+ motion_bucket_id: int = 127,
34
+ fps_id: int = 6,
35
+ version: str = "svd_xt",
36
+ cond_aug: float = 0.02,
37
+ decoding_t: int = 3, # Number of frames decoded at a time! This eats most VRAM. Reduce if necessary.
38
+ device: str = "cuda",
39
+ output_folder: str = "outputs",
40
+ ):
41
+ if image.mode == "RGBA":
42
+ image = image.convert("RGB")
43
+
44
+ if(randomize_seed):
45
+ seed = random.randint(0, max_64_bit_int)
46
+ generator = torch.manual_seed(seed)
47
+
48
+ os.makedirs(output_folder, exist_ok=True)
49
+ base_count = len(glob(os.path.join(output_folder, "*.mp4")))
50
+ video_path = os.path.join(output_folder, f"{base_count:06d}.mp4")
51
 
52
+ frames = pipe(image, decode_chunk_size=decoding_t, generator=generator, motion_bucket_id=motion_bucket_id, noise_aug_strength=0.1, num_frames=25).frames[0]
53
+ export_to_video(frames, video_path, fps=fps_id)
54
+ torch.manual_seed(seed)
55
+
56
+ return video_path, seed
57
 
58
+ def resize_image(image, output_size=(1024, 576)):
59
+ # Calculate aspect ratios
60
+ target_aspect = output_size[0] / output_size[1] # Aspect ratio of the desired size
61
+ image_aspect = image.width / image.height # Aspect ratio of the original image
62
 
63
+ # Resize then crop if the original image is larger
64
+ if image_aspect > target_aspect:
65
+ # Resize the image to match the target height, maintaining aspect ratio
66
+ new_height = output_size[1]
67
+ new_width = int(new_height * image_aspect)
68
+ resized_image = image.resize((new_width, new_height), Image.LANCZOS)
69
+ # Calculate coordinates for cropping
70
+ left = (new_width - output_size[0]) / 2
71
+ top = 0
72
+ right = (new_width + output_size[0]) / 2
73
+ bottom = output_size[1]
74
+ else:
75
+ # Resize the image to match the target width, maintaining aspect ratio
76
+ new_width = output_size[0]
77
+ new_height = int(new_width / image_aspect)
78
+ resized_image = image.resize((new_width, new_height), Image.LANCZOS)
79
+ # Calculate coordinates for cropping
80
+ left = 0
81
+ top = (new_height - output_size[1]) / 2
82
+ right = output_size[0]
83
+ bottom = (new_height + output_size[1]) / 2
84
 
85
+ # Crop the image
86
+ cropped_image = resized_image.crop((left, top, right, bottom))
87
+ return cropped_image
 
 
88
 
89
+ with gr.Blocks() as demo:
90
+ gr.Markdown('''# Community demo for Stable Video Diffusion - Img2Vid - XT ([model](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt), [paper](https://stability.ai/research/stable-video-diffusion-scaling-latent-video-diffusion-models-to-large-datasets), [stability's ui waitlist](https://stability.ai/contact))
91
+ #### Research release ([_non-commercial_](https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt/blob/main/LICENSE)): generate `4s` vid from a single image at (`25 frames` at `6 fps`). this demo uses [🧨 diffusers for low VRAM and fast generation](https://huggingface.co/docs/diffusers/main/en/using-diffusers/svd).
92
+ ''')
93
+ with gr.Row():
94
+ with gr.Column():
95
+ image = gr.Image(label="Upload your image", type="pil")
96
+ generate_btn = gr.Button("Generate")
97
+ video = gr.Video()
98
+ with gr.Accordion("Advanced options", open=False):
99
+ seed = gr.Slider(label="Seed", value=42, randomize=True, minimum=0, maximum=max_64_bit_int, step=1)
100
+ randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
101
+ motion_bucket_id = gr.Slider(label="Motion bucket id", info="Controls how much motion to add/remove from the image", value=127, minimum=1, maximum=255)
102
+ fps_id = gr.Slider(label="Frames per second", info="The length of your video in seconds will be 25/fps", value=6, minimum=5, maximum=30)
103
+
104
+ image.upload(fn=resize_image, inputs=image, outputs=image, queue=False)
105
+ generate_btn.click(fn=sample, inputs=[image, seed, randomize_seed, motion_bucket_id, fps_id], outputs=[video, seed], api_name="video")
106
 
107
+ if __name__ == "__main__":
108
+ #demo.queue(max_size=20, api_open=False)
109
+ demo.launch(share=True, show_api=False)