Research map / Image generation / Autoregressive and latent generation
Multi-Scale Representation Alignment for Visual Autoregressive Modeling with Mixture of Experts
Nuoyan Zhou, Zhijun Tu, Lei Yu, Kun Cheng, jie hu, Nannan Wang, Xinghao Chen
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 scale-aware token-routed MoE (STMoE) (ablation) | VAR-d16 · ImageNet 256 (VAR-d16, 100 epochs) | FID: 4.10 -> 2.99 | Improved |
| add semantic guidance (SG) alignment (ablation) | VAR-d16 · ImageNet 256 (VAR-d16, 100 epochs) | FID: 4.10 -> 3.59 | Improved |
| add semantic guidance on top of STMoE (ablation) | VAR-d16 + STMoE · ImageNet 256 (VAR-d16, 100 epochs) | FID: 2.99 -> 2.65 | Improved |
| add STMoE on top of semantic guidance (ablation) | VAR-d16 + SG · ImageNet 256 (VAR-d16, 100 epochs) | FID: 3.59 -> 2.65 | Improved |
| use token-routed STMoE instead of scale-routed SMoE (ablation) | VAR-d16 · ImageNet 256 (VAR-d16, 100 epochs) | FID: 3.11 -> 2.99 | Improved |
| add MoEEC routing (no scale awareness) (ablation) | VAR-d16 · ImageNet 256 (VAR-d16, 100 epochs) | FID: 4.10 -> 4.06 | No clear effect |
| add MoE++ routing (ablation) | VAR-d16 · ImageNet 256 (VAR-d16, 100 epochs) | FID: 4.10 -> 3.39 | Improved |
| use STMoE instead of SMoE for expert load balance (ablation) | VAR-d16 · VAR-d16 routed-token load | expert load variance: 6.62 -> 2.41 | Improved |
| align with Mocov3 features (ablation) | VAR-d16 · ImageNet 256 (VAR-d16, 100 epochs) | FID: 4.10 -> 4.11 | No clear effect |
| align with DINOv3 instead of DINOv2 (ablation) | VAR-d16 + SG · ImageNet 256 (VAR-d16, 100 epochs) | FID: 3.81 -> 3.59 | Improved |
| align aggregated intermediate-scale features (SG-M) instead of only the final aggregated feature (SG-Last) (ablation) | VAR-d16 + SG · ImageNet 256 (VAR-d16, 100 epochs) | FID: 3.81 -> 3.59 | Improved |
| train MEPA-d16 for 200 instead of 100 epochs | MEPA-d16 · ImageNet 256x256 | FID: 2.65 -> 2.32 | 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
- Markovian Scale Prediction: A New Era of Visual Autoregressive Generation (CVPR 2026 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · similarity 0.88)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · similarity 0.87)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.87)
- Data-efficient Large Vision Models through Sequential Autoregression (ICML 2024 · similarity 0.86)
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.
