is 124 frames the limit?
When I increase video length to more than 5seconds (124 frames), the model behaves weirdly. I already tried 8 steps (anything lower, and it just doesn't work)
Also, is it me, or does viggle-pruned-int8-convrot perform much slower than the h3-ref2va-int8-convrot model? Gen time is almost double, even with sparse attention
the model was trained on 124 frames yes and thats where it performs best, if you want to do more than 5 second video you should download my experimental branch it has a single looping Ksampler workflow wich is easier or an advanced long video loop it can do much longer videos by feeding the next video, the previous 22 frames and the ref image. i will merge to main once i test it more
https://github.com/Saganaki22/ComfyUI-Viggle-Animate-H3/tree/exp
this is one of the examples
On another note.
I think you are supposed to use image edit and match ref_character with 1 frame of the person in the video. Then use the edited image as a reference image in the workflow.
Otherwise, feeding a random ref character image just doesn't work
yes everything is explained in the readme