Research map / Image generation / Autoregressive and latent generation
Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis
Peng Zheng, Junke Wang, Yi Chang, Yizhou Yu, Rui Ma, Zuxuan Wu
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 |
|---|---|---|---|
| condition MAR on discrete tokens (ablation) | MAR · ImageNet 256x256 | gFID: reduces gFID by up to 0.2 points, IS up about 10 points | Improved |
| add discrete token conditioning to MAR-H (fewer AR steps) (ablation) | MAR-H · ImageNet 256x256 | sec/img: 1.678 -> 0.316 | Improved |
| use a stronger discrete AR model (RAR-XXL vs RAR-B) to produce conditions (ablation) | DisCon-L · ImageNet 256x256 | gFID: 1.87 -> 1.38 | Improved |
| condition on weaker discrete AR models (RAR-B, RAR-L) (ablation) | DisCon · ImageNet 256x256 | gFID: 1.74 -> 1.74 (RAR-L with DisCon-B) | No clear effect |
| reduce AR steps from 256 to 64 in MAR vs DisCon at 16 steps (ablation) | MAR vs DisCon · ImageNet 256x256 | gFID | Improved |
| apply CFG to continuous token generation (ablation) | DisCon · ImageNet 256x256 | gFID | Hurt |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.90)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.90)
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.90)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.89)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.89)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · 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.
