Research map / Image generation / Autoregressive and latent generation
Autoregressive Image Generation with Masked Bit Modeling
Qihang Yu, Qihao Liu, Ju He, Xinyang Zhang, Yang Liu, Liang-Chieh Chen, Xi Chen
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 |
|---|---|---|---|
| use masked bit modeling head instead of linear head (16-bit) (ablation) | BAR-B · ImageNet-256 (w/ CFG) | gFID: 1.63 -> 1.19 | Improved |
| use masked bit modeling head instead of bit head (10-bit) (ablation) | BAR-B · ImageNet-256 (w/ CFG) | gFID: 2.63 -> 1.48 | Improved |
| use MBM head to enable 32-bit codebook where linear head is infeasible (ablation) | BAR-B · ImageNet-256 | feasibility / gFID: 1.37 | Improved |
| scale codebook from 10 to 64 bits with bit head (ablation) | BAR-B · ImageNet-256 (w/ CFG) | gFID: 2.63 -> 10.97 | Hurt |
| scale codebook from 16 to 64 bits with MBM head (ablation) | BAR-B · ImageNet-256 (w/ CFG) | gFID: 1.19 -> 1.67 | Mixed |
| use uniform instead of arccos masking ratio sampling (16-bit) (ablation) | BAR-B · ImageNet-256 (w/ CFG) | FID: 1.46 -> 1.22 | Improved |
| use logit-normal instead of uniform masking ratio sampling (16-bit) (ablation) | BAR-B · ImageNet-256 (w/ CFG) | FID: 1.22 -> 1.19 | No clear effect |
| enlarge prediction head from 3x1024 to 6x2048 (32-bit, w/o CFG) (ablation) | BAR-B · ImageNet-256 (w/o CFG) | FID: 2.68 -> 2.10 | Improved |
| enlarge prediction head under CFG (12-bit, 3x1024 to 6x2048) (ablation) | BAR-B · ImageNet-256 (w/ CFG) | FID: 1.26 -> 1.27 | No clear effect |
| increase unmasking steps from 2 to 3 or more (16-bit) (ablation) | BAR-B · ImageNet-256 (w/o CFG) | FID: 3.73 -> 1.95 | Improved |
| use back-loading unmasking schedule [2, 2, 5, 7] instead of uniform [4, 4, 4, 4] (ablation) | BAR-B · ImageNet-256 | FID: 1.19 -> 1.15 (w/ CFG); 1.68 -> 1.81 (w/o CFG) | Mixed |
| use token shuffling with patch size 2 (64 tokens, 64 bits) (ablation) | BAR-B · ImageNet-256 (w/ CFG) | FID and images/sec: 1.19 -> 1.35; 24.9 -> 150.3 images/sec | Mixed |
Closest papers by meaning
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.90)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · similarity 0.89)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.89)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · similarity 0.89)
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.
