Research map / Image generation / Autoregressive and latent generation
STAT: Soft Tail-dropping for Adaptive Visual Tokenization
Zeyuan Chen, Kai Zhang, Zhuowen Tu, Yuanjun Xiong
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 adaptive token allocation instead of a fixed token count (220) (ablation) | STAT tokenizer · ImageNet-1k class-conditional generation (LlamaGen-XL) | gFID: 2.73 -> 2.36 | Improved |
| use adaptive token allocation instead of a fixed token count (220) (ablation) | STAT tokenizer · ImageNet-1k val 256x256 reconstruction | rFID: 1.15 -> 1.15 | No clear effect |
| use adaptive token allocation instead of a fixed token count (220) (ablation) | STAT tokenizer · ImageNet-1k val 256x256 reconstruction | PSNR: 20.35 -> 20.22 | Hurt |
| use per-token probabilistic (Bernoulli) dropping instead of hard tail-drop (ablation) | STAT tokenizer · ImageNet-1k class-conditional generation (LlamaGen-XL) | gFID: 2.67 -> 2.36 | Improved |
| use per-token probabilistic (Bernoulli) dropping instead of hard tail-drop (ablation) | STAT tokenizer · ImageNet-1k val 256x256 reconstruction | PSNR: 20.30 -> 20.22 | Hurt |
| sample the EoS threshold during AR training instead of a fixed threshold (ablation) | LlamaGen-STAT-XL · ImageNet-1k class-conditional generation (LlamaGen-XL) | gFID: 2.49 -> 2.36 | Improved |
| add content-adaptive loss (ablation) | STAT tokenizer · ImageNet-1k reconstruction | token allocation | Improved |
| add decreasing importance loss (ablation) | STAT tokenizer · ImageNet-1k reconstruction | tail-dropping pattern | Improved |
| lower the inference threshold from 0.5 to 0.01 | STAT · ImageNet-1k val 256x256 reconstruction | rFID: 1.15 -> 0.99 | Improved |
| add 10 extra tokens at inference on top of threshold 0.01 | STAT · ImageNet-1k val 256x256 reconstruction | rFID: 0.99 -> 0.88 | Improved |
| use expected token count instead of threshold 0.5 at inference (ablation) | STAT · ImageNet-1k val 256x256 reconstruction | rFID: 1.15 -> 1.20 | No clear effect |
| replace the LlamaGen VQ tokenizer with STAT in text-conditional AR | vanilla AR model · GenEval | Overall: 0.32 -> 0.45 | Improved |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior (ICLR 2025 · similarity 0.88)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.88)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.88)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 2025 · similarity 0.88)
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.
