Research map / Image generation / Autoregressive and latent generation
Multi-Scale Local Speculative Decoding for Image Generation
Elia Peruzzo, Guillaume Sautière, Amirhossein Habibian
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 |
|---|---|---|---|
| replace LANTERN with MuLo-SD on Tar-7B | Tar-7B 1024p · MS-COCO 5k val, 1024p | speedup: 1.45x -> 5.33x | Improved |
| replace LANTERN with MuLo-SD on Tar-1.5B | Tar-1.5B 1024p · MS-COCO 5k val, 1024p | GenEval: 75.4 -> 76.8 | Improved |
| use MuLo-SD instead of EAGLE-2 on Tar-1.5B | Tar-1.5B 512p · MS-COCO 5k val, 512p | speedup: 0.72x -> 1.94x | Improved |
| apply LANTERN to the stronger Tar models instead of LlamaGen | LANTERN · MS-COCO 5k val, 512p | speedup: 1.59x -> 1.08x | Hurt |
| add MuLo-SD to LlamaGen-XL | LlamaGen-XL 512p · MS-COCO 5k val, 512p | speedup: 1.40x vs 1.59x LANTERN | Mixed |
| add local rejection and resampling instead of raster-scan rejection (ablation) | MuLo-SD · GenEval vs speedup, 512p | GenEval | Improved |
| add neighborhood expansion to local verification (ablation) | MuLo-SD · GenEval vs speedup, 512p | GenEval | Improved |
| add codebook-neighbor probability pooling (ablation) | MuLo-SD · GenEval vs speedup, 512p | GenEval | Improved |
| add parallel decoding (ZipAR drafting, island resampling) (ablation) | MuLo-SD · MS-COCO 5k val | latency | Improved |
| train up/down-samplers with pixel-space MSE and LPIPS instead of token classification loss (ablation) | MuLo-SD upsampler · visual fidelity | perceptual quality | Improved |
| add PatchGAN adversarial loss to upsampler training (ablation) | MuLo-SD upsampler · visual fidelity | perceptual quality | Improved |
| use off-the-shelf pixel-space upsampler instead of learned latent upsampler (ablation) | MuLo-SD upsampler · GenEval vs speedup, 512p | GenEval | Hurt |
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.
- + ZipAR-16 — ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025) · better on 5 of 5 shared results
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
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.90)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.89)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.88)
- Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding (ICLR 2025 · similarity 0.87)
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · similarity 0.87)
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.
