Desm0nt

> Gamechanging or not, it's unlikely to catch on if hobbyists can't play around with it, so the fact that this technique requires a workstation class gpu is a huge...

@hopibel Jupiter notebook on vast.ai works the same way as in colab =) Textual inversion is good and really very comfortable compare to this solution, but it actually can't add...

Same issue

> and how long wait? im sick of dalle2 waiting already Training models of this scale is not an easy task, not fast at all, and most importantly, not cheap....

Even though Tiled VAE works with SDXL - it still has a problem that SD 1.5 didn't have, specifically a weird dot/grid pattern 

Lora training (even not a low VRAM, fp8-quanto with 23.5 also good enough) for windows users would be amazing (SimpleTuner not working even on WSL, only rented linux env)

> Flux LoRa training capability on a 24gb card is rather limited I run rank-16 batch 1 gradient accumulation 3 on single 4090 in 1024*1024 in fp8-quanto in an attempt...

Completely agree. Technicaly it is the real next SD model (instead of SD 3) from SD authors, and it's deserve to be run on really comfortable UI (instead of Comfy)

> What is needed to get this done? Am i wrong in assuming it is not too big of a task seeing how quickly comfy supported it? It's not a...

> > > > This if FLUX1-dev checkpoint with included VAE, maybe it can help https://huggingface.co/Comfy-Org/flux1-dev/tree/main > > > > > > > > > Waaau!! Great thank to u...