Research map / Image generation / Autoregressive and latent generation
CODA: Repurposing Continuous VAEs for Discrete Tokenization
Zeyu Liu, Zanlin Ni, Yeguo Hua, Xin Deng, Xiao Ma, Cheng Zhong, Gao Huang
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 residual quantization (10 levels) to plain VQ on a continuous VAE (ablation) | CODA · ImageNet 256x256, 16x | rFID: 41.52 -> 14.23 | Improved |
| add attention-based sparse quantization (ablation) | CODA · ImageNet 256x256, 16x | rFID: 14.23 -> 7.06; utilization 51% -> 100% | Improved |
| adapt the VAE with LoRA during discretization (ablation) | CODA · ImageNet 256x256, 16x | rFID: 7.06 -> 1.34 | Improved |
| tune LoRA on decoder versus encoder (ablation) | CODA · ImageNet 256x256, 16x | rFID: encoder 7.06 -> 5.75; decoder 7.06 -> 1.59 | Mixed |
| scale codebook size from 1024 to 65536 (ablation) | CODA · ImageNet 256x256, 16x | rFID: 2.07 -> 1.34 | Improved |
| normalize query and key in attention quantizer (ablation) | CODA · ImageNet 256x256, 16x | rFID: 1.87 -> 1.34 | Improved |
| use RMSNorm instead of LayerNorm (ablation) | CODA · ImageNet 256x256, 16x | rFID: 1.37 -> 1.34 | No clear effect |
| use CODA tokenizer instead of VQGAN training | discrete tokenizer · ImageNet 256x256 | training iterations and utilization: 300K -> 50K iterations; 5.9% -> 100% utilization | Improved |
| use CODA tokenizer in MaskGIT instead of original tokenizer | MaskGIT · ImageNet 256x256 generation | FID: 4.92 -> 3.17 | 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.
- MaskGIT — MaskGIT: Masked Generative Image Transformer (CVPR 2022) · 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
- UniTok: a Unified Tokenizer for Visual Generation and Understanding (NeurIPS 2025 · similarity 0.89)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 2025 · similarity 0.89)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.89)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.88)
- Hybrid Token Compression for Vision-Language Models (CVPR 2026 · 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.
