@mckenna don't know, haven't looked into that, but, judging based on what i've been able to train on a 4090, cloud may be the only real option for video training on this. training on a 4090 with just a few images rides right at 23.4gb, 5090 would have better luck, and of course, the professional cards with 48+ gb vram. even 2 3090's with nvlink.

@mckenna so you are in the same boat as me, pushing the limits of 24gb vram trying to train this thing.
in musubi, it does allow the fp16 models, probably won't take much to train against the fp8 distilled, that should help, but I haven't gone down that path yet, still working on the over-fitting issue, backgrounds and camera position fit WAAAAY before anything else.

@mckenna so a bit of some info, this is all using musubi tuner.
yesterday, i got my first video training set going, started with simple bj videos, I have to limit to 256x256 @ 49 frames otherwise oom.
but thanks to something I learned back in hyv1, as long as you have some matching images for these videos, that quality is ALOT better then just 256x256 clips alone, it has a high res reinforcement of that scene (768x768)
but, 16 clips, 16 high res stills of those clips, 2 hours of training on a 4090. it does reaaaaly well, looking at it, could of let it train for another hour, those 2 hours were a total of 100 epochs (3200 steps).
today, I spent some time.. probably way to much time, building up a more generalized set, hoping to use as a base for all my future training, the idea is this set has the basic understand that hyv1 had originally (hyv1 could do a lot even without lora's)
thats my goal with this set.
so a selection of clips for standard adult content, 36 in total, with 3 high res reinforcements pulled from each clip. so 36 videos, 108 images (set in different sets or oom)
i have clips running on repeat of 3 so the images don't over-fit waay before motion is trained.

the biggest slowdown for me at the moment, is rank/dim of 8, might be able to push 16, but 32 oom'ed on videos no matter what I did (unless i wanted 1 second clips)

musubi is nice, its smart with clip training, only loads 1 clip at a time, so as long as 1 clip fits, all fit.
one of the things i did, to get more frames into training, was take a 3 second 24fps clip, dropped to 16 fps, then saved it again as 24 fps, that shrinks 3 seconds into 2, and yes, it is a fast clip, but, part of why my last night was just test, it did exactly what I wanted it to do.
i can generate at 16fps now and its not all in weird slowmo, because 24fps on these clips is technically fast forward.
once I see how this current training comes out, ill probably start dropping some motion trained lora's for 1.5

@mckenna still working on character training setup, but in the process of figuring out best settings for that, nailed styles.
trying to make a jinx lora, haven't nailed jinx yet, but nailed arcane style, lol