THUDM

/

CogVideoX1.5-5B-I2V

like 74

Follow

Knowledge Engineering Group (KEG) & Data Mining at Tsinghua University 1.43k

I have the same error, using torch and torchao nightly and diffusers, transformers compiled from source 11/17/2024

how do you solve this problem

huggingface/diffusers#9877

https://github.com/zRzRzRzRzRzRzR/diffusers/tree/cogvideox1.1-5b
Try installing diffusers from here, as the PR has not been merged yet.

huggingface/diffusers#9877

https://github.com/zRzRzRzRzRzRzR/diffusers/tree/cogvideox1.1-5b
Try installing diffusers from here, as the PR has not been merged yet.

I have already downloaded version 0.32.0.dev0 of diffusers following the link you provided, but the issue still persists.

This fixed my error:pip uninstall diffusers and then pip install git+https://github.com/zRzRzRzRzRzRzR/[email protected] -U

This fixed my error:pip uninstall diffusers and then pip install git+https://github.com/zRzRzRzRzRzRzR/[email protected] -U

its work for me! Thank you!

Please pay attention to our current update, I have modified two configurations to allow I2V to work at multiple resolutions, remember to update the configuration file.

Please pay attention to our current update, I have modified two configurations to allow I2V to work at multiple resolutions, remember to update the configuration file.

Do you mean the hypes width and height in cli_demo.py?

No, it's the latest commit to update this repo(Huggingface Not Github). cli_demo doesn't need to be modified.

No, it's the latest commit to update this repo(Huggingface Not Github). cli_demo doesn't need to be modified.

Could you please explain why you changed “sample_height” and “sample_width” from (96, 170) to (300, 300)? For the default input size (1360, 768), it is downsampled 8 times by the VAE to become (170, 96). The previous settings seemed correct, so I’m quite confused. Do I also need to adjust the input width and height accordingly? I look forward to your response.

I would also like to ask for your advice. When testing image2video with the x1.5-5b model, I encountered an issue where the generated 5-second video suddenly has a camera jump. Specifically, the first few seconds look fine, but then there’s an abrupt scene change in the middle, making the rest of the video quite unpleasant. Could this be due to some parameter not being set correctly? I follow the default setting by the cli_demo.

If it is the last second, this is possible, especially when the resolution is not 1360 * 768 (I can also reproduce a similar situation). Whether each video has a similar problem？

If it is the last second, this is possible, especially when the resolution is not 1360 * 768 (I can also reproduce a similar situation). Whether each video has a similar problem？

Yes, 80% of the videos might have this issue, but not just in the last second. For example, in a 5-second video, the last 2 or even 3 seconds might exhibit this problem.

I updated diffuser version to your latest 0.32.0.dev0, but I still got an error related to diffusers

"File "/opt/conda/lib/python3.11/site-packages/diffusers-0.32.0.dev0-py3.11.egg/diffusers/pipelines/pipeline_loading_utils.py", line 756, in _fetch_class_library_tuple
library = not_compiled_module.module.split(".")[0]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'bool' object has no attribute 'module'. Did you mean: 'mod'?"

Any idea how to handle this issue?

huggingface/diffusers#9877

https://github.com/zRzRzRzRzRzRzR/diffusers/tree/cogvideox1.1-5b
Try installing diffusers from here, as the PR has not been merged yet.

this worked

Sometimes, you will encounter an HTTP problem. And I tried an off-line way to solve this problem:

download the source file from: https://github.com/zRzRzRzRzRzRzR/diffusers; you will get a zip file "diffusers-main.zip".
directly run: pip install diffusers-main.zip.

Now it can be installed directly from diffusers 0.32.0.