How to Train the Flux-Fill-dev Model? Issues with Latent and Mask Concatenation

[yyyyyxie]

Hi,
I want to training the flux-fill-dev, but I’m unsure where to start. Specifically, I always seem to get the latent representations using VAE, but I’m unsure how to concatenate the latent and the mask according to the channel concatenation method described in flux-fill-dev.

[bghira]

i dont support ip adapter or inpainting training

[yyyyyxie]

Thank you for your help. How I can save the quantized weight files obtained using your framework, specifically in the safetensors format?

[bghira]

do you mean the int8 or fp8 versions created at startup? currently you cannot.