4 upvotes
StepX-OPD merges capabilities and speed in text-to-image diffusion
What is StepX-OPD: Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models?
StepX-OPD trains a single diffusion model that can follow instructions, render embedded text and align with human preferences, while still generating images in just a few sampling steps. It does this by distilling knowledge from multiple specialist teachers directly during on-policy training.
Who it is for
Anyone building or researching deployable text-to-image services can use it to get a more capable, faster model without stitching together separate stages.
How to try it
Open the paper on Hugging Face: https://huggingface.co/papers/2611.04386