Research map / Image generation / Autoregressive and latent generation
GigaTok: Scaling Visual Tokenizers to 3 Billion Parameters for Autoregressive Image Generation
Tianwei Xiong, Jun Hao Liew, Zilong Huang, Jiashi Feng, Xihui Liu
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 semantic regularization with frozen DINOv2 features to tokenizer training (ablation) | GigaTok vanilla tokenizer · ImageNet AR probing | gFID / validation loss: gFID consistently decreasing; validation loss relatively constrained | Improved |
| scale decoder instead of encoder (S-B vs B-S) (ablation) | GigaTok · ImageNet AR probing | gFID: 6.56 -> 5.65 | Improved |
| also scale encoder (S-L to B-L) (ablation) | GigaTok · ImageNet AR probing | gFID: 5.19 -> 4.82 | Improved |
| use DINOv2-B instead of CLIP as semantic encoder (ablation) | GigaTok S-B · ImageNet AR probing | gFID: 6.35 -> 5.55 | Improved |
| use DINOv2-B instead of SigLIP as semantic encoder (ablation) | GigaTok S-B · ImageNet AR probing | gFID: 6.20 -> 5.55 | Improved |
| align at decoder layer 4 instead of layer 2 (ablation) | GigaTok S-S · ImageNet AR probing | gFID / linear probing accuracy: gFID 6.26 -> 6.07; Lin Acc. 63.4 -> 58.6 | Mixed |
| raise semantic regularization weight from 0.25 to 1.00 (ablation) | GigaTok S-S · ImageNet AR probing | rFID / linear probing accuracy: rFID 1.28 -> 1.38; Lin Acc. 57.0 -> 62.5 | Mixed |
| add entropy loss for 2.9B XL-XXL tokenizer (ablation) | GigaTok XL-XXL · ImageNet training | codebook usage / convergence | Improved |
| use 1D Q-Former tokenizer instead of 2D ViT tokenizer (ablation) | GigaTok with semantic regularization · ImageNet AR probing | rFID / linear probing accuracy / gFID trend | Improved |
| add DINO discriminator | GigaTok B-L · ImageNet 256x256 | rFID / gFID: rFID 0.81 -> 0.51; gFID 3.26 -> 3.33 | Mixed |
| scale tokenizer from S-S to XL-XXL | GigaTok · ImageNet 256x256 (LlamaGen-B) | gFID: 4.05 -> 3.15 | Improved |
| scale tokenizer from B-L to XL-XXL | GigaTok · ImageNet 256x256 (LlamaGen-XXL) | gFID: 2.03 -> 1.98 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ImageFolder — ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learnings from Scaling Visual Tokenizers for Reconstruction and Generation (ICML 2025 · similarity 0.92)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.91)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.91)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025 · similarity 0.90)
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.90)
- Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · 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.
