Research map / Image generation / Autoregressive and latent generation
Draft-and-Revise: Effective Image Generation with Contextual RQ-Transformer
Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, Wook-Shin Han
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 classifier-free guidance only in the revise phase (ablation) | Contextual RQ-Transformer 821M · ImageNet | FID: 5.78 -> 3.45 | Improved |
| apply guidance in draft phase too (ablation) | Contextual RQ-Transformer 821M · ImageNet | FID: 3.45 -> 8.90 | Hurt |
| apply guidance in draft phase too (ablation) | Contextual RQ-Transformer 821M · ImageNet | Recall: 0.52 -> 0.33 | Hurt |
| increase T_draft from 8 to 64 (ablation) | Contextual RQ-Transformer · ImageNet | FID: 5.41 -> 3.45 | Improved |
| increase T_draft from 8 to 64 (ablation) | Contextual RQ-Transformer · ImageNet | s/sample: 0.03 -> 0.10 | Hurt |
| increase number of parameters from 821M to 1.4B | Contextual RQ-Transformer · ImageNet | FID: 3.45 -> 3.41 | No clear effect |
| increase parameters from 371M to 821M | Contextual RQ-Transformer · ImageNet | FID: 5.45 -> 3.45 | Improved |
| increase T_revise (ablation) | Contextual RQ-Transformer 821M · ImageNet | FID | Hurt |
| revise phase after draft phase (ablation) | Contextual RQ-Transformer · ImageNet | FID: 15.32 -> 3.45 | Improved |
| Top-C confidence-based infilling instead of random in draft (T_draft=64) (ablation) | Contextual RQ-Transformer 821M · ImageNet | FID: 3.45 -> 27.88 | Hurt |
| Top-C-50% instead of random in draft (T_draft=64) (ablation) | Contextual RQ-Transformer 821M · ImageNet | FID: 3.45 -> 6.84 | Hurt |
| Top-C-50% instead of random (T_draft=8) (ablation) | Contextual RQ-Transformer 821M · ImageNet | FID: 6.28 -> 6.13 | 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
- Toward Accurate Image Generation via Dynamic Generative Image Transformer (TPAMI 2026 · similarity 0.88)
- A Spark of Vision-Language Intelligence: 2-Dimensional Autoregressive Transformer for Efficient Finegrained Image Generation (ICLR 2025 · similarity 0.85)
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 2025 · similarity 0.85)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.85)
- Masked Generative Nested Transformers with Decode Time Scaling (ICML 2025 · similarity 0.85)
- JetFormer: An autoregressive generative model of raw images and text (ICLR 2025 · similarity 0.85)
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.
