RECON paper: reusing REPA's projection head at inference makes DiT training 25x faster
udmrzn · x · 2026-10-17
The arXiv paper "Representation Conditioning for Generation" (Saeed Razavi et al.) proposes RECON, improving semantic injection in diffusion transformers:
- Problem: REPA discards its MLP projection head after training, so later DiT layers supervised only by velocity loss aren't guaranteed access to semantics.
- Method: During training, feed pretrained vision-encoder tokens into later DiT layers; at inference, reuse the MLP head as a proxy for the encoder and project its outputs into those layers.
- Results: Zero extra parameters, 63x faster convergence than SiT and 25x faster than REPA, better ImageNet FID, and high quality retained with fewer sampling steps.
- Code is open-sourced.
More from Research
- Plotly and Mila launch DataAnalyticsBench: 43 tasks, 28 models on a 20M-row warehouse — Mila_Quebec · 2026-10-17
- Terry Tao's Four Types of Mathematization: Heavy AI Use Needs Strong Human Understanding — ravithejads · 2026-10-17
- Princeton's LabOS, Featured in Nature: An AI Co-Scientist for Hypothesis and Experiment Design — MengdiWang10 · 2026-10-17
- 'Tip-of-the-tongue' queries: a blind spot for modern agentic retrieval — bclavie · 2026-10-17
- Tip-of-the-tongue queries emerge as a key test of agentic retrieval systems — ctjlewis · 2026-10-17
- micro1 launches LegalReasonBench to test models on attorney-authored legal matters — douwekiela · 2026-10-17