LoRA Training
Chroma seems to somewhat work with Flux LoRAs. I tried with some of my character LoRAs I made for dev, and they give acceptable results.
However, they throw tons of text encoder and Unet errors.
Is there any way to train for chroma yet, especially looking at Fluxgym. Since it's based on schnell, maybe it can be done?
Chroma is so good, I would abandon Flux any moment if I could re-train my LoRAs.
I've been told you may get good discussions and support on the discord on this topic (but I've not been there yet).
AFAIK, diffusion-pipe and ai-toolkit already integrate Chroma.
Awesome, I will look into ai-toolkit! Thanks for the hint!
Awesome, I will look into ai-toolkit! Thanks for the hint!
You're very much welcome!
If I'm not mistaken, diffusion-pipe allows masked training and multi-captioning, like OneTrainer, but I haven't used it yet so it has to be confirmed. It's in command-line, but seems quite simple, with just a config file (and examples are provided).