Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

THUDM/CogVideoX-5b-I2V is producing jibberish output #517

Open
1 of 2 tasks
nitinmukesh opened this issue Nov 18, 2024 · 5 comments
Open
1 of 2 tasks

THUDM/CogVideoX-5b-I2V is producing jibberish output #517

nitinmukesh opened this issue Nov 18, 2024 · 5 comments

Comments

@nitinmukesh
Copy link

System Info / 系統信息

diffusers 0.32.0.dev0
torch 2.5.1+cu121
torchvision 0.20.1+cu121
python 3.11

Information / 问题信息

  • The official example scripts / 官方的示例脚本
  • My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

python inference/cli_demo.py --prompt "A young girl with sun-kissed hair and sparkling blue eyes stands in a lush, sunlit garden, her face radiant with a genuine smile that lights up her entire being. She wears a soft, floral dress that complements the vibrant blooms around her. As she tilts her head slightly, the sunlight catches the gentle curve of her smile, highlighting her joyful expression. Her hands are gently clasped in front of her, adding to the serene and happy atmosphere. The background is a tapestry of colorful flowers and greenery, enhancing the warmth and beauty of her smile." --model_path THUDM/CogVideoX-5b-I2V --generate_type "i2v" --num_frames 48 --image_or_video_path image.png --width 720 --height 480

output-i2v.mp4

Expected behavior / 期待表现

I tried the same image here and output was fine

https://huggingface.co/spaces/THUDM/CogVideoX-5B-Space

@a-r-r-o-w
Copy link

a-r-r-o-w commented Nov 18, 2024

num_frames must be 81 or 161. I see it mentioned in the docs here: https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox.

Looks like the table rendering is broken, so will fix that in a follow-up PR

@nitinmukesh
Copy link
Author

@a-r-r-o-w

I am using --model_path THUDM/CogVideoX-5b-I2V, I think it supports less than 49 frames.
THUDM/CogVideoX1.5-5b-I2V supports 81 frames.

Is my understanding wrong?

@a-r-r-o-w
Copy link

Oh really sorry, I misread the model id. 49 frames is indeed correct here, but I see that you have 48 frames specified in the command. The I2V model is very sensitive to this and that's probably why you're seeing this tile patterns

@nitinmukesh
Copy link
Author

So what do you suggest. Please if you can guide on how to do inference using v1 i2v.

@a-r-r-o-w
Copy link

Try running inference with height=480, width=720 and num_frames=49. I don't get any artifacts running the pure Diffusers example. I will take a look at the file in question soon as well to see if there are any bugs

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants