Please refrain from baking in vae and text encoders. We all have these large bits downloaded already. There is no need to bloat your checkpoints with them. My SSD begs of you.

unless there are serious concerns about security, and as long as you're experimenting, could everyone please use Q5 and Q8 GGUF format? because they run a lot faster and on more cards, take up less space, download from the internet faster, etc. at least while folks are experimenting. when checkpoint creators strike gold and create a perfect model, then they can be 15GB and 30GB in size and that would make the most sense.

Thanks for the trimmed down version. I've done some initial testing and this one works well with appearance loras. Can't wait for future iterations. Anatomy can be cleaned up for now using any of the existing loras here.

I like it. Works pretty well across the board. Seems to work best with a higher guidance of around 4 but doesn't seem too picky about the sampler/scheduler.

Are there any plans to upload an FP16 model? I much prefer the look of your finetune compared to the base Flux Dev and would love to train LoRAs on it specifically, but LoRA training seems to only be supported with FP16 models at the moment