Research map / Image generation / Autoregressive and latent generation
Scalable Image Tokenization with Index Backpropagation Quantization
Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, Limin Wang
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 |
|---|---|---|---|
| replace VQ with IBQ (index backpropagation to all codes) (ablation) | VQGAN tokenizer (16,384 codes, dim 256) · ImageNet val reconstruction | rFID: 3.98 -> 1.67 rFID; usage 5.3% -> 98% | Improved |
| add double quantization loss (ablation) | IBQ · ImageNet val reconstruction | rFID: 1.67 -> 1.55 | Improved |
| use a deeper tokenizer (4 ResBlocks instead of 2) (ablation) | IBQ + double quantization · ImageNet val reconstruction | rFID: 1.55 -> 1.37 | Improved |
| increase codebook size from 1,024 to 262,144 (ablation) | IBQ · ImageNet val reconstruction | rFID: 2.24 -> 1.00 | Improved |
| increase codebook size to 262,144 (ablation) | IBQ · ImageNet val reconstruction | codebook usage: 96% -> 84% | Hurt |
| increase code dimension from 32 to 256 (ablation) | IBQ · ImageNet val reconstruction | rFID: 2.04 -> 1.37 | Improved |
| increase code dimension from 64 to 128 to 256 (ablation) | IBQ · ImageNet val reconstruction | rFID: 1.39 -> 1.38 -> 1.37 | No clear effect |
| increase ResBlocks in encoder/decoder from 1 to 4 (ablation) | IBQ · ImageNet val reconstruction | rFID: 1.80 -> 1.37 | Improved |
| increase code dimension from 8 to 256 in standard VQ (LlamaGen) | LlamaGen tokenizer · ImageNet val reconstruction | rFID: 2.19 -> 9.21; usage 97% -> 0.29% | Hurt |
| use IBQ instead of LFQ tokenizer (ablation) | vanilla AR transformer 342M, 16,384 codes · ImageNet 256x256 generation | gFID: 3.40 -> 2.88 | Improved |
| scale AR model from 342M to 2.1B with IBQ tokenizer | IBQ-B to IBQ-XXL · ImageNet 256x256 generation | gFID: 2.88 -> 2.05 | Improved |
| replace hard-index backprop with soft quantization (weighted code average) in training (ablation) | IBQ · ImageNet val reconstruction | rFID: 16.17 vs 4.03 (same epochs); soft train/hard infer 233.17 | Hurt |
Closest papers by meaning
- Scaling the Codebook Size of VQ-GAN to 100,000 with a Utilization Rate of 99% (NeurIPS 2024 · similarity 0.90)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.89)
- LG-VQ: Language-Guided Codebook Learning (NeurIPS 2024 · similarity 0.89)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.88)
- Codebook Transfer with Part-of-Speech for Vector-Quantized Image Modeling (CVPR 2024 · similarity 0.88)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · 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.
