Great work! I wish I had the computer power and skills to train AI on 3D meshes. I have to settle for trying to do AI enhancements on 3D renders to try to squeeze some photo-realism from them, Really cool work there.

Hey, thank you.

As long as your GPU has 8 gigs of VRAM to spare, you should still be able to perform some local training tasks at varying degrees of efficiency.

I started testing these things out with just a 8 GB RTX 2080 way back in 2022, when unoptimized embeddings were the only tools available for training on checkpoints with meager hardware so I can say for sure that it's definitely gotten much better now with more training options publicly available with less requirements for beefy hardware. Since I don't really trust any of the online / cloud training alternatives, I wouldn't personally recommend them...but those also are another option.

I'm working on a lappy with an integrated NVIDIA card, so I think I'm kind of screwed there. Still really keen on your work, though.

In that case an external solution (eGPU) should help out 'on paper'...but last time I checked such an option came with its own can of worms since it was quite fiddly to assign the correct GPU for training purposes (it's been a while, I haven't looked into the current builds for such edge cases so things might have now improved) and even then it will definitively be less powerful than a dedicated internal card when it comes to VRAM management or it/sec values.

On a side note, I've just started delving deeper into AI-gen animation, in trying to figure out the technical bits of motion modules and how lora/lyco blocks can affect movement and composition over various checkpoint versions. It's nowhere as robust as that newly released Sora generator, but the folks working on the motion modules are also holding their own especially since they aren't back by billion/trillion dollar corps. So lots to explore on that front as well.

But in any case, please do look forward to the new stuff I'm going upload soon, been making silent headway on a lot of stuff and I'll let the pictures (and videos) speak for themselves :)

I'll look for your upcoming work. you're light years ahead of me on both the hardware and tech savvy. For now, I'm locked into starting with DAZ 3D renders for consistency and creative control since I usually work in image series. For pure AI, I just don't see a way around that.

Agreed, consistency is right now AI's weakest suit...mutated fingers might come a close second place.

I dunno if you're already using control net, but selective masking + controlnet refining in img2img will allow you a much tighter control structure even with using just Daz screenshots as reference image, as opposed to full iray renders. Adetailer is also another strong recommendation, and if you're adventurous (or fixated) enough you will be able to train your own search-&-replace models for outfits/footwear/backgrounds/whatever if you side-grade to a cloud training service or some additional hardware option just for training the models out.

I'm putting some finishing touches on two custom AI models trained from my own assets right now: the bodypaint freebie that I share over here and the bodystraps lingerie harness for G8F product that's also sold over here. Once I distill out the most viable block weights, the generations from those models should give you a good idea of how the pure AI-gens can compare with some 3D asset which you can load and manipulate inside DS, without having to use a daz render/preview as primer for the final image.