search

PKU-YuanGroup / Open-Sora-Plan

This project aim to reproduce Sora (Open AI T2V model), we wish the open source community contribute to this project.
MIT License
11.18k stars 996 forks source link

video low quality for 256x256 #330

Open quantumiracle opened 1 month ago

quantumiracle commented 1 month ago

Hi,

I try to generate 65x256x256 video with the provided 65x512x512 model, however the quality is low as shown below: opensora256 should the DiT framework generalizable for different image resolutions? or if I miss some details in code?

LinB203 commented 1 month ago

This is normal, absolute positional coding has very limited expandability.

LinB203 commented 1 month ago

We will replace it into RoPE in the next version.

quantumiracle commented 1 month ago

Thanks, could you also provide a low-resolution model like 256x256 for research purpose, if you have one.