Decoding Register Tokens Preserves Global Scene Details
Pretrained vision transformers learn structured image representations spanning spatial patch tokens and non-spatial tokens.
ProofPaper ↗
Key points
- Recent analyses suggest that non-spatial register tokens encode global scene properties such as illumination, blur, and style.
- We investigate whether retaining register features improves decoding when the complete spatial patch representation is already available.
- Appearance probes, paired reconstruction-error analysis, and fixed-decoder register substitutions examine which global image properties are accessible from registers and how their values affect reconstructed pixels.
- On ImageNet-1K, our ViT-B decoders achieve mean reconstruction FIDs of 0.472 with DINOv2 and 0.442 with DINOv3, reductions of 9.5% and 5.2% relative to their respective patch-only baselines.
Sources (1)
- [1]Decoding Register Tokens Preserves Global Scene DetailsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 09:46 AM
Pretrained vision transformers learn structured image representations spanning spatial patch tokens and non-spatial tokens.
Recent analyses suggest that non-spatial register tokens encode global scene properties such as illumination, blur, and style.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 9, 2026Qwen/Qwen-Image-2.1-Turbo
- Oct 8, 2026Dissecting Representation Structure in Vision Transformers: A Rigorous Architectural Study
- Oct 8, 2026MCL: Meta Convolution Layer
- Oct 8, 2026One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts
- Oct 8, 2026LEGO: A Lifting-Free Approach for Exocentric-to-Egocentric Video Generation
- Oct 7, 2026On the Necessity of Attention-FFN Split in Vision Transformers