Add diffusers example
#20
by
patrickvonplaten
- opened
README.md
CHANGED
|
@@ -16,6 +16,61 @@ zeroscope_v2_XL uses 15.3gb of vram when rendering 30 frames at 1024x576
|
|
| 16 |
2. Replace the respective files in the 'stable-diffusion-webui\models\ModelScope\t2v' directory.
|
| 17 |
### Upscaling recommendations
|
| 18 |
For upscaling, it's recommended to use the 1111 extension. It works best at 1024x576 with a denoise strength between 0.66 and 0.85. Remember to use the same prompt that was used to generate the original clip.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
### Known issues
|
| 20 |
Rendering at lower resolutions or fewer than 24 frames could lead to suboptimal outputs. <br />
|
| 21 |
|
|
|
|
| 16 |
2. Replace the respective files in the 'stable-diffusion-webui\models\ModelScope\t2v' directory.
|
| 17 |
### Upscaling recommendations
|
| 18 |
For upscaling, it's recommended to use the 1111 extension. It works best at 1024x576 with a denoise strength between 0.66 and 0.85. Remember to use the same prompt that was used to generate the original clip.
|
| 19 |
+
|
| 20 |
+
### Usage in 🧨 Diffusers
|
| 21 |
+
|
| 22 |
+
Let's first install the libraries required:
|
| 23 |
+
|
| 24 |
+
```bash
|
| 25 |
+
$ pip install git+https://github.com/huggingface/diffusers.git
|
| 26 |
+
$ pip install transformers accelerate torch
|
| 27 |
+
```
|
| 28 |
+
|
| 29 |
+
Now, let's first generate a low resolution video using [cerspense/zeroscope_v2_576w](https://huggingface.co/cerspense/zeroscope_v2_576w).
|
| 30 |
+
|
| 31 |
+
```py
|
| 32 |
+
import torch
|
| 33 |
+
from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
|
| 34 |
+
from diffusers.utils import export_to_video
|
| 35 |
+
|
| 36 |
+
pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16)
|
| 37 |
+
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
|
| 38 |
+
pipe.enable_model_cpu_offload()
|
| 39 |
+
pipe.enable_vae_slicing()
|
| 40 |
+
|
| 41 |
+
prompt = "Darth Vader is surfing on waves"
|
| 42 |
+
video_frames = pipe(prompt, num_inference_steps=40, height=320, width=576, num_frames=36).frames
|
| 43 |
+
video_path = export_to_video(video_frames)
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
Next, we can upscale it using [cerspense/zeroscope_v2_XL](https://huggingface.co/cerspense/zeroscope_v2_XL).
|
| 47 |
+
|
| 48 |
+
```py
|
| 49 |
+
pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_XL", torch_dtype=torch.float16)
|
| 50 |
+
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
|
| 51 |
+
pipe.enable_model_cpu_offload()
|
| 52 |
+
pipe.enable_vae_slicing()
|
| 53 |
+
|
| 54 |
+
video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames]
|
| 55 |
+
|
| 56 |
+
video_frames = pipe(prompt, video=video, strength=0.6).frames
|
| 57 |
+
video_path = export_to_video(video_frames, output_video_path="/home/patrick/videos/video_1024_darth_vader_36.mp4")
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
Here are some results:
|
| 61 |
+
|
| 62 |
+
<table>
|
| 63 |
+
<tr>
|
| 64 |
+
Darth vader is surfing on waves.
|
| 65 |
+
<br>
|
| 66 |
+
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/darth_vader_36_1024.gif"
|
| 67 |
+
alt="Darth vader surfing in waves."
|
| 68 |
+
style="width: 576;" />
|
| 69 |
+
</center></td>
|
| 70 |
+
</tr>
|
| 71 |
+
</table>
|
| 72 |
+
|
| 73 |
+
|
| 74 |
### Known issues
|
| 75 |
Rendering at lower resolutions or fewer than 24 frames could lead to suboptimal outputs. <br />
|
| 76 |
|