kpsss34/ModularX-Portrait-77token

introduction

I continued distill & training this model as an extension of black-forest-labs/FLUX.1-Krea-dev, replacing T5xxl with T5large. I then started training it on a specialized dataset. The results are quite good and comparable to the full model. It achieves a high accuracy of around 80–90%, while using only 24GB of VRAM (using the diffusers pipeline, not ComfyUI). Its only limitation is that it can handle prompts of up to 77 tokens only.

The pipeline for using the model is still under improvement and development. I hope it will be good enough for people who don’t want to use long prompts.