zai-org
how to reproduce the demo's generated video quality?
testing the model on my machine ,the quality of the generated video is much worse than demo show case.
Do you using GLM-4 or GPT-4 to enhance the prompt? maybe can check web_demo.py
If you using the demo prompt. it should worked using with CogVideoXDDIMSchedule with 50 step in fp16
In today's various generative models, prompt polishing is very important for the generated results.
I can only maybe get to like 50-70% of quality to their examples, even with the same prompts. Are the examples really of this model or of their next "Pro" model?
It is this model and we will update pro model demo video when it is release. did you using with 50 step DDIM sampler and gs 6
I have the same problem. The quality of the reproduced example seems to be only around 50-70% in exactly same condition.
May there is a little performance loss when converting from sat version to hf version. You can try sat version for inference or wait for our better model soon.