Advice on Memory management - CUDA Out of Memory

[BarnabasTakacs]

Excellent work, congratulations.

I have run into CUDA Out of Memory errors when trying with 480p
I need to run the code on an older configuration.
What is your suggestion to change parameters?

Thank you.
Barnabas Takacs

[Chenxi-Song]

Hi Barnabas,

Thank you for your interest and for trying our code.

Unfortunately, although the 480p and 720p versions use different weights, their weight sizes are essentially the same. Therefore, the runtime memory requirement is also very similar. In our setting, the Wan-based model requires around 69 GB of GPU memory.

If you are using a lower-memory GPU, reducing the resolution alone may not be sufficient. A more practical option may be to switch to video generation backbones with quantization, pruning, or other memory-efficient variants. You may want to check existing video models and choose one that can fit within your hardware constraints.

[BarnabasTakacs]

Thank your for such a quick response.

My specific application calls for a human-centric approach.
Is it the same situation for the LongCat-Video-Avatar?

What is the recommended version you are running the inference?

Thank you
Barnabas Takacs

[Chenxi-Song]

Thank you for the follow-up.

As far as I know, LongCat-Video-Avatar may still require a similar amount of GPU memory, since it is also based on a large video generation model. I have not extensively tested the Avatar-specific setting, so I cannot give a very precise recommendation for that case.

In our experiments, we mainly run inference with the standard LongCat-Video model rather than the Avatar version. If memory is the main bottleneck, I would suggest considering more efficient or distilled video models, such as Helios or other recent distilled variants, which may be more suitable for lower-memory devices.