Research map / Image generation / Autoregressive and latent generation
Autoregressive Image Generation Using Residual Quantization
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 residual quantization with depth 4 instead of a single VQ code at 8×8 resolution (ablation) | RQ-VAE · ImageNet validation reconstructions | rFID: 17.95 -> 4.73 | Helped |
| Increase residual quantization depth from 2 to 4 (ablation) | RQ-VAE · ImageNet validation reconstructions | rFID: 10.77 -> 4.73 | Helped |
| Increase residual quantization depth beyond 4 (ablation) | RQ-VAE · ImageNet validation reconstructions | rFID: 4.73 -> 2.69 -> 1.83 | Helped |
| Increase VQ codebook size from 16,384 to 131,072 at 8×8×1 (ablation) | VQ-GAN · ImageNet validation reconstructions | rFID: 17.95 -> 17.09 | No clear effect |
| Replace the shared codebook with four non-shared codebooks (ablation) | RQ-VAE · ImageNet validation reconstruction | rFID: 4.73 -> 5.73 | Hurt |
| Train the RQ-VAE for 50 epochs rather than the shorter training setup (ablation) | RQ-VAE and RQ-Transformer · ImageNet class-conditioned generation | FID: 11.56 -> 8.71 | Helped |
| Apply stochastic sampling during RQ-Transformer training (ablation) | RQ-Transformer, 821M · ImageNet class-conditioned generation | FID: 14.06 -> 13.24 | Helped |
| Use soft labeling without stochastic sampling (ablation) | RQ-Transformer, 821M · ImageNet class-conditioned generation | FID: 14.06 -> 14.87 | Hurt |
| Combine stochastic sampling and soft labeling (ablation) | RQ-Transformer, 821M · ImageNet class-conditioned generation | FID: 14.06 -> 13.11 | Helped |
| Use the full RQ-Transformer rather than the variant without stochastic sampling and soft labeling (ablation) | RQ-Transformer, 821M · ImageNet class-conditioned generation | FID: 14.06 -> 13.11 | Helped |
| Increase model parameters from 1.4B to 3.8B (ablation) | RQ-Transformer with 50-epoch RQ-VAE · ImageNet class-conditioned generation | FID: 8.71 -> 7.55 | Helped |
| Use a shorter image-code sequence with RQ-VAE and RQ-Transformer | RQ-Transformer versus VQ-GAN · ImageNet class-conditioned generation sampling | speedup: 4.1×; 5.6×; 7.3× | Helped |
Closest papers by meaning
- A Spark of Vision-Language Intelligence: 2-Dimensional Autoregressive Transformer for Efficient Finegrained Image Generation (ICLR 2025 · similarity 0.89)
- Visual Autoregressive Modeling for Image Super-Resolution (ICML 2025 · similarity 0.88)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.87)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.87)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · similarity 0.87)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.86)
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.