{"slug": "representation-by-design-in-generation-cross-view-class-token-alignment-in", "title": "Representation by Design in Generation: Cross-View Class-Token Alignment in Diffusion Transformers", "summary": "A new arXiv paper (2609.36348v1) reports that adding cross-view class-token alignment to diffusion transformer training improves ImageNet linear-probing accuracy by 9.4% with the class token and 10.1% with mean-pooled patch tokens versus a matched two-view baseline, while frozen-backbone VOC2012 segmentation gains 3.6 mIoU. The method, which aligns each student class-token representation with a stop-gradient EMA-teacher target from a second independently noised observation, maintains comparable ImageNet generation FID and cuts text-to-image FID from 2.52 to 2.37 at matched checkpoints. The authors conclude that semantic representation can be optimized as a first-class capability of diffusion pretraining alongside generation rather than remaining a by-product of synthesis.", "body_md": "arXiv:2609.36348v1 Announce Type: new \nAbstract: Generative and representation learning remain asymmetrically connected: semantic representations are used to improve diffusion generation, whereas the models' own representations are often treated as a by-product of synthesis. We ask whether diffusion models can instead be trained to learn substantially stronger semantic representations without sacrificing generation quality. SelfFlow takes a step in this direction by introducing self-supervised patch alignment into flow matching, but its main gains remain in faster convergence and improved generation. Inspired by DINO and iBOT, we extend this framework with cross-view class-token alignment to further strengthen semantic representations. Specifically, we form two independently noised, dual-timestep observations of each image and align each student class-token representation with the stop-gradient EMA-teacher target from the other observation. This objective is optimized jointly with the inherited flow-matching and local patch objectives. Notably, although the additional objective acts only on the class token, it strengthens both class-token and patch representations. Compared with a matched two-view baseline, ImageNet linear-probing accuracy improves by 9.4\\% using the class token and 10.1\\% using mean-pooled patch tokens, while frozen-backbone VOC2012 segmentation improves by 3.6 mIoU. These representation gains are achieved while maintaining comparable ImageNet generation FID. In text-to-image training, the same objective also improves generation FID, reducing it from 2.52 to 2.37 at matched checkpoints. Our results show that representation need not remain a by-product of generation or merely a tool for improving it: it can be directly optimized as a first-class capability of diffusion pretraining alongside generation.", "url": "https://wpnews.pro/news/representation-by-design-in-generation-cross-view-class-token-alignment-in", "canonical_source": "https://arxiv.org/abs/2609.36348", "published_at": "2026-09-30 04:00:00+00:00", "updated_at": "2026-09-30 04:21:30.148126+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "generative-ai", "computer-vision", "ai-research"], "entities": ["SelfFlow", "DINO", "iBOT", "ImageNet", "VOC2012", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/representation-by-design-in-generation-cross-view-class-token-alignment-in", "markdown": "https://wpnews.pro/news/representation-by-design-in-generation-cross-view-class-token-alignment-in.md", "text": "https://wpnews.pro/news/representation-by-design-in-generation-cross-view-class-token-alignment-in.txt", "jsonld": "https://wpnews.pro/news/representation-by-design-in-generation-cross-view-class-token-alignment-in.jsonld"}}