Research map / Image generation / Autoregressive and latent generation
Drift-AR: Single-Step Visual Autoregressive Generation via Anti-Symmetric Drifting
zhen zou, Xiaoxiao Ma, Mingde Yao, Jie Huang, Linjiang Huang, Feng Zhao
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 |
|---|---|---|---|
| use entropy-parameterized variance instead of constant variance (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.72 -> 1.57 | Improved |
| use entropy-parameterized variance instead of constant variance (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | IS: 289.3 -> 298.1 | Improved |
| use annealed alpha(t) instead of fixed alpha=0.5 (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.69 -> 1.57 | Improved |
| add repulsion term (anti-symmetric kernel objective) (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.69 -> 1.57 | Improved |
| add entropy-based early stopping in speculative decoding (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.62 -> 1.57 | Improved |
| freeze AR prior in Phase II (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.67 -> 1.57 | Improved |
| use drifting decoder instead of diffusion decoder at 1 step (ablation) | TransDiff-H · ImageNet 256x256 (TransDiff-H) 1 step | FID: 14.72 -> 1.57 | Improved |
| use drifting decoder instead of DMD at 1 step (ablation) | TransDiff-H · ImageNet 256x256 (TransDiff-H) 1 step | FID: 2.93 -> 1.57 | Improved |
| sigma_max 1.0 instead of 0.5 (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.57 -> 1.64 | Hurt |
| sigma_max 0.1 instead of 0.5 (ablation) | Drift-AR · ImageNet 256x256 (TransDiff-H) | FID: 1.57 -> 1.68 | Hurt |
| apply Drift-AR to NextStep-1 | NextStep-1 · GenEval / MJHQ-30K | FID: 6.71 -> 6.66 | Improved |
| apply DMD to NextStep-1 | NextStep-1 · GenEval / MJHQ-30K | FID: 6.71 -> 8.33 | Hurt |
Closest papers by meaning
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.90)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.88)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · 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.
