Research map / Image generation / Autoregressive and latent generation
Test-Time Registers as Global Priors for Tokenized Image Generation
Cheng-Yao Hong, Yifan Wang, Yuewei Lin, Chenyu You
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| inject RegToken prior into HCT decoding | HCT · ImageNet HCT decoding (DINOv2-L/14 prior) | FID-5k: 21.2 -> 20.1 | Improved |
| inject RegToken prior into HCT decoding | HCT · ImageNet HCT decoding (DINOv2-L/14 prior) | SigLIP: 3.5 -> 3.9 | Improved |
| inject RegToken prior (OpenCLIP backbone) | HCT · ImageNet HCT decoding (OpenCLIP-L/14 prior) | FID-5k: 20.8 -> 20.3 | Improved |
| inject random prior instead of none | HCT · ImageNet HCT decoding (DINOv2-L/14 prior) | FID-5k: 21.2 -> 22.5 | Hurt |
| naive test-time register prior instead of no prior | HCT · ImageNet HCT decoding (DINOv2-L/14 prior) | FID-5k: 21.2 -> 21.5 | Hurt |
| use RegToken prior in token optimization | HCT · Test-time optimization efficiency | Steps@tau: 74 -> 52 | Improved |
| use prior from same image rather than random image (ablation) | RegToken · Prior source (DINOv2-L/14) | FID-5k: 21.1 -> 20.1 | Improved |
| add NFN layer selection (ablation) | RegToken · Component ablation (DINOv2-L/14 + HCT) | FID-5k: 21.1 -> 20.9 | Improved |
| add TokenRank head gating (ablation) | RegToken · Component ablation (DINOv2-L/14 + HCT) | FID-5k: 21.3 -> 20.8 | Improved |
| add projection-and-conservation interpolation (ablation) | RegToken · Component ablation (DINOv2-L/14 + HCT) | FID-5k: 20.8 -> 20.3 | Improved |
| vary number of register neurons (16 or 64 vs 32) (ablation) | RegToken · Hyperparameter ablation (DINOv2-L/14 + HCT) | FID-5k: 20.3 -> 20.7 | Hurt |
| change register scale (1.0 or 4.0 vs 2.0) (ablation) | RegToken · Hyperparameter ablation (DINOv2-L/14 + HCT) | FID-5k: 20.3 -> 21.0 | Hurt |
Closest papers by meaning
- Register and [CLS] tokens induce a decoupling of local and global features in large ViTs (NeurIPS 2025 · similarity 0.91)
- Vision Transformers Don't Need Trained Registers (NeurIPS 2025 · similarity 0.90)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.89)
- Attention (as Discrete-Time Markov) Chains (NeurIPS 2025 · similarity 0.87)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.86)
- RATS! Patches Talk Through Registers: Emergent Parts in Register Attention Transformers (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
