Research map / Image generation / Autoregressive and latent generation
FastHybrid: Accelerating Hybrid Autoregressive Image Generation with Lookahead and Guided Decoding
Zhengguo Jiang, Fang Zhang, Yongxiang Hua, Bocheng Li, Wentao Zhang, Linli Xu
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 FastHybrid instead of halving AR steps (64 to 32) | MAR-L · ImageNet 256x256 | FID: 2.05 -> 1.90 at similar speedup (1.97x vs 1.92x) | Improved |
| use FastHybrid instead of halving AR steps | MAR-H · ImageNet 256x256 | FID: 1.92 -> 1.70 | Improved |
| apply FastHybrid to MAR-B | MAR-B · ImageNet 256x256 | FID: 2.32 -> 2.43 | Hurt |
| apply FastHybrid to MAR-H | MAR-H · ImageNet 256x256 | running time (s): 35.7 -> 21.0 | Improved |
| add guided diffusion sampling after rejection filtering (RG-0.8 vs R-0.8) (ablation) | FastHybrid · ImageNet (10k, MAR-B) | FID: 4.94 -> 4.84 | Improved |
| raise semantic similarity rejection threshold from 0.0 to 0.8 (ablation) | FastHybrid · ImageNet (10k, MAR-B) | FID: 5.60 -> 4.84 | Improved |
| raise rejection threshold (cost) (ablation) | FastHybrid · ImageNet (10k, MAR-B) | time (s): 7.47 -> 10.02 | Hurt |
| reduce diffusion steps from 50 to 30 without guidance (ablation) | MAR-D50 · ImageNet (10k) | FID: 4.75 -> 5.04 | Hurt |
| use add-noise-then-denoise (inverse) instead of guided sampling (ablation) | FastHybrid · ImageNet (10k) | FID: 4.84 -> 4.87 / 5.13 | Hurt |
| use decreasing (cosine) guidance weight instead of increasing (Linear-up) (ablation) | FastHybrid · ImageNet (10k) | FID: 5.13 -> 4.84 | Improved |
| use cosine schedule instead of linear-down or square schedule (ablation) | FastHybrid · ImageNet (10k) | FID: 4.99 / 5.09 -> 4.84 | Improved |
Closest papers by meaning
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 2025 · similarity 0.90)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.90)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.90)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 2025 · similarity 0.89)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · 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.
