RECON speeds up diffusion training 25x via representation conditioning

The RECON paper proposes representation conditioning for diffusion Transformers, reusing REPA projection heads at inference to train 25x faster than REPA. A related paper from ANU's Liang Zheng team reaches similar conclusions.

2026-10-17 ~ 2026-10-18 · 2 related posts

• RECON paper: reusing REPA's projection head at inference makes DiT training 25x faster — udmrzn · 2026-10-17
• RECON paper: representation conditioning beats alignment, 25x faster than REPA — udmrzn · 2026-10-18