frontier models
A Discarded Projection Head Comes Back at Inference
RECON routes visual-encoder representations into later diffusion-transformer layers and reuses the alignment head after the clean training image disappears.
Summary
RECON routes visual-encoder representations into later diffusion-transformer layers and reuses the alignment head after the clean training image disappears.
Representation-aligned diffusion training usually teaches an intermediate layer with a vision encoder, then throws away the projection head. RECON keeps that pathway useful: during training it conditions later layers on encoder tokens; during inference the retained head stands in for the unavailable clean-image representation. The authors say the change adds no parameters and converges 63 times faster than SiT and 25 times faster than REPA in their comparisons. Those ratios depend on the reported setups, but the design closes a clear gap between early semantic supervision and later generation layers.
Why it matters
RECON routes visual-encoder representations into later diffusion-transformer layers and reuses the alignment head after the clean training image disappears.
Limits and context
No additional limitation was separately recorded.
Key claims
RECON routes visual-encoder representations into later diffusion-transformer layers and reuses the alignment head after the clean training image disappears.
Evidence: source-2026-10-15-005
Sources
- arXiv preprint 2610.20608, original v1arXiv / submitting research authors · primary research
Corrections
: Publication note: This October 15 edition was prepared before its scheduled 9:00 AM America/Detroit reveal, but publication was delayed by a production build failure. The original reporting and image-generation timestamps are preserved.