Research map / Architectures and efficiency / Mixture of experts
Prism-MoE: Efficient Dense-to-MoE Conversion for Visual Autoregressive Generation
Ying Li, Zefang Wang, Zhaode Wang, Zhiwen Chen, chengfei lv, Huan Wang
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 trajectory-consistent initialization (ablation) | Taylor importance pruning init · VAR-d16 ImageNet, 75% active, no FT | FID: 27.5 -> 12.6 | Improved |
| add distillation fine-tuning at 37.5% active (ablation) | trajectory-consistent init · VAR-d16 ImageNet | FID: 12.6 -> 9.00 | Improved |
| add normalized cosine router (ablation) | distillation fine-tuned MoE · VAR-d16 ImageNet, 37.5% active | FID: 9.00 -> 6.04 | Improved |
| add spatially-aware routing loss (ablation) | cosine router MoE · VAR-d16 ImageNet, 37.5% active | FID: 6.04 -> 4.26 | Improved |
| fine-tune longer with standard load-balancing loss (epoch 5 to 20) (ablation) | VAR-d20 MoE, 37.5% active · ImageNet-1K C2I | FID: 7.05 -> 10.7 | Hurt |
| replace linear softmax router + LB loss with Prism-MoE fine-tuning | VAR-d20 MoE, 37.5% active · ImageNet-1K C2I | FID: 10.7 -> 2.86 | Improved |
| replace linear softmax router + LB loss with Prism-MoE fine-tuning | LlamaGen-XL MoE, 37.5% active · MS-COCO T2I | FID: 31.5 -> 26.0 | Improved |
| increase calibration set size from 100 to 200 (ablation) | Prism-MoE init · VAR-d20 FID | FID: 6.04 -> 5.03 | Improved |
| increase calibration set size from 200 to 300 (ablation) | Prism-MoE init · VAR-d20 FID | FID: 5.03 -> 4.96 | No clear effect |
| increase ridge multiplier from 1.0 to 2.0 (ablation) | Prism-MoE init · VAR-d20 FID | FID: 5.03 -> 5.28 | Hurt |
| use Prism-MoE init at 75% active instead of dense | Semanticist · Semanticist | FID: 2.49 -> 2.99 | Hurt |
| use Prism-MoE (37.5% active) instead of dense | VAR-d20 · Inference latency | latency: 45.5 ms -> 37.2 ms | Improved |
Closest papers by meaning
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.88)
- Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE (NeurIPS 2026 · similarity 0.87)
- Data-efficient Large Vision Models through Sequential Autoregression (ICML 2024 · similarity 0.87)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.87)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.87)
- Mixture of Tokens: Continuous MoE through Cross-Example Aggregation (NeurIPS 2024 · 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.
