Research map / Image generation / Autoregressive and latent generation
VibeToken: Scaling 1D Image Tokenizers and Autoregressive Models for Dynamic Resolution Generations
Maitreya Patel, Jingtao Li, Weiming Zhuang, Yezhou Yang, Lingjuan Lv
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 |
|---|---|---|---|
| add dynamic grid position embedding in place of learnable RoPE (ablation) | TiTok-S-128 VAE tokenizer · VAE small-scale study, 1024x1024 | rFID: 1.544 -> 1.707 (256), 93.49 -> 131.20 (1024) | Mixed |
| add adaptive patch embedding on top of dynamic grid embedding (ablation) | TiTok-S-128 VAE tokenizer · VAE small-scale study, 1024x1024 | rFID: 131.20 -> 5.38 | Improved |
| add plain RoPE to the TiTok baseline (ablation) | TiTok-S-128 VAE tokenizer · VAE small-scale study, 256x256 | rFID: 1.862 -> 2.934 | Hurt |
| add learnable RoPE to the TiTok baseline (ablation) | TiTok-S-128 VAE tokenizer · VAE small-scale study, 256x256 | rFID: 1.862 -> 1.544 | Improved |
| use 64 tokens instead of 256 for the AR generator (ablation) | VibeToken-Gen GPT-B · ImageNet 256 gFID with cfg | gFID: 9.81 -> 8.42 | Improved |
| use 128 tokens instead of 256 for the AR generator (ablation) | VibeToken-Gen GPT-B · ImageNet 256 gFID with cfg | gFID: 9.81 -> 9.02 | Improved |
| train as a mixed-resolution generalist with 64 tokens (ablation) | VibeToken-Gen GPT-B · ImageNet 256 gFID | gFID: w/o cfg 33.15 -> 31.14, w/ cfg 8.42 -> 9.37 | Mixed |
| train as a mixed-resolution generalist with 128 tokens (ablation) | VibeToken-Gen GPT-B · ImageNet 256 gFID | gFID: w/o cfg 37.68 -> 39.50, w/ cfg 9.02 -> 10.58 | Hurt |
| replace LlamaGen (576 tokens, fixed resolution) with VibeToken-Gen (64 tokens, generalist) | GPT-B AR generator · ImageNet 256 gFID | gFID: w/o cfg 33.44 -> 31.14, w/ cfg 7.15 -> 9.37 | Mixed |
| replace diffusion NiT with VibeToken-Gen-XXL for 1024x1024 generation | class-conditional generator · ImageNet 1024x1024 | gFID: 5.87 -> 3.54 | Improved |
| scale VibeToken-Gen from B to XXL | VibeToken-Gen · Low-resolution average gFID | gFID: 9.14 -> 5.63 | Improved |
| use VibeToken-LL (large encoder) instead of VibeToken-SL | VibeToken tokenizer · Arbitrary-resolution reconstruction | rFID: 4.21 -> 3.60 | Improved |
Closest papers by meaning
- Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · similarity 0.90)
- Native-Resolution Image Synthesis (NeurIPS 2025 · similarity 0.90)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.89)
- BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 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.
