https://jethrocsau.github.io/tags/representation-alignment/Representation AlignmentHugo Blox Builder (https://hugoblox.com)en-usWed, 09 Sep 2026 00:00:00 +0000https://jethrocsau.github.io/media/icon_hu502917898142073417.pnghttps://jethrocsau.github.io/tags/representation-alignment/https://jethrocsau.github.io/post/autorepa/Wed, 09 Sep 2026 00:00:00 +0000https://jethrocsau.github.io/post/autorepa/<p>Representation alignment is usually trained as a static, always-on auxiliary loss for diffusion transformers. This project (with Tianyu Zhao) asks whether that assumption is necessary — studying the question in three stages on a SiT-B/2 backbone trained on ImageNet-100 at 256px latent diffusion: validating the baseline ordering and layer placement, testing hard early-stop schedules for the projection term, and evaluating long-horizon cosine annealing over 250k steps.</p> <p><strong><a href="https://jethrocsau.github.io/auto-repa-blog/">Read the full write-up →</a></strong></p> <p><a href="https://github.com/zhaotianyu0702/Auto-REPA" target="_blank" rel="noopener">Source Code</a></p>