Spaces:
Running
on
L40S
Cannot reimplement the results from demo
Thanks for your great work! i've successefully deployment your project locally, i know you mention that there're some limitations on your hf app demo, so according to your notes, i edited the correspongding part, such as using the full graph, but i think there's still some wired frame jumps in the results, like this:
any suggestion to fix this?ah, this one is not "frame jump", it is a natural camera zoom in and zoom out in the reference video. and just for this speaker.
- other speakers will be fine.
- for this speaker, you can delete these edges manually. since you have setup in your local. you may have the logs, and there is a path like [45, 46, 47, 48, ...., 765] for generating this video. you could locate and delete the edges with camear zoom in. for example frame 47 -> 48, by graph.detelte(47, 48).
thanks for your replying, i'll try it. and one more question, i just realised that the frame interpolation in your code is FiLM as you mentioned for faster processing, but i suppose the model ACInterp you proposed in the paper would achieve much better result right? is there any plan you will open that model?
yes, but I plan to update that part latter such as next month. that code base is based on moore-animate anyone, I haven't clean up it.