A hybrid model
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/tree/main
Would it be possible to do 10 Eros with this configuration?
Currently deciding whether to use your model on this one and would rather not choose, as both provide functionality the other does not.
I'm thinking about doing that for future versions yeah, or maybe this other linked version of the hybrid I saw which is the ref model but with the fl deltas on it. I just need to see confirmation that it's fully accurate on video and image reference. The fl model can already kind of do reference outputs which gives a skewed interpretation that there merges work, but I haven't compared to the full ref model. It might not actually do video reference restyling as accurately.
https://huggingface.co/xmarre/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024-ComfyUI/tree/main
Thanks, I'll download that one and try it out. Good luck with the merge.
I was going to bring this up too. Lol.
It's pretty insane how many potential combinations are possible with H3.
Yes, I'm a bit skeptical about the quality (as in better) of merged. I feel they would have done this from the outset if that was the case. More interesting was a reddit post where someone did multiple tests between base and reference, both doing I2V, T2V and Ref2V and there was hardly any difference apart from some audio ones. I've just been using Ref for I2V anyway to be honest.
ref for i2v works good. It adjusts what you prompt differently from the start image but otherwise the reference=startframe fairly reliable.