Very interested Merge
I really like your experiment with the Vace merge.
Vace is such a good concept, and its a shame not more models do this kind of style of model.
I was testing with the model, and noticed with anime outpainting that lines and movements are excellent, just colors are not great.
When compared besides the skyreelsV2 T2V Vace merge it looks as capable if not for the colors.
Have you found any good combination of samplers, schedulers or lora to get it to shine?
Hey Thanks for your interest. I think I ended up preferring euler/eulerA, more high steps without loras and the lightx for the last 50% for speed boost.
In the end though I think bytedance's new model bernini is bascially what I had envisioned here (its made by the same people who made seedance and I really feel it their 'reasearch' model that is based on the multimodal inputs that that model is based on. It should be implimented into main ComfyUI soon.