"perframe_ae" was not specified when calling with resolution 512 from gradio_app.py

[ngc-shj]

Great work!

When I checked the code in run.sh and run_mp.sh, the "perframe_ae" option was specified when the resolution was other than 256, i.e. 512 and 1024. However, "perframe_ae" is not set at resolution 512 because only the values from the config file are applied when invoking from gradio_app.py.

I think adding "perframe_ae: True" to inference_512_v1.0.yaml is a good idea. What do you think?

[Doubiiu]

yes. i was aware of this when releasing code. But actually perframe_ae=True won't bring any benefits for DynamiCrafter512 model in terms of saving GPU memory based on my own attempt. Sure you can add that in the config.

[ngc-shj]

Thank you for you replying.

In my environment, RTX 4090 24GB, CPU offload occurred in DynamiCrafter512 model. Immediately after starting gradio_app.py, 12.8GB of VRAM was used, and after that, memory was used up to 37.2GB (vram 23.6 + mem 13.6). Setting perframe_ae=True did not use more than 12.8GB of VRAM. The time required to generate a video has also been reduced from 120.51 seconds to 53.26 seconds.

[Doubiiu]

wow Thanks! I will add that into the config of 512 model.

[ngc-shj]

FYI. For, RTX 4090 Laptop GPU (16GB)

• 1024 model: CUDA OOM occurred
• 512 model: 12.6GB、102.69 sec perframe_ae: True
• 256 model: 11.8GB、46.48 sec perframe_ae: True

[Doubiiu]

Thanks for your contribution!🙏 @ngc-shj