Research map / Image generation / Autoregressive and latent generation
Mirai: Autoregressive Visual Generation Needs Foresight
Yonghao Yu, Lang Huang, Zerun Wang, Runyi Li, Toshihiko Yamasaki
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 Mirai-I foresight alignment to LlamaGen-B | LlamaGen-B · ImageNet 256x256, 300 epochs | FID: 5.34 -> 4.34 | Improved |
| add Mirai-I to LlamaGen-XL | LlamaGen-XL · ImageNet 256x256, 300 epochs | FID: 3.16 -> 2.59 | Improved |
| inject foresight at output level (multi-token prediction) instead of internal alignment (ablation) | LlamaGen-B · ImageNet 256x256, 80 epochs | FID: 6.36 -> 7.28 | Hurt |
| align internal representations to foresight in 2D grid rather than 1D scan (ablation) | LlamaGen-B · ImageNet 256x256, 80 epochs | FID: 6.20 -> 5.22 | Improved |
| increase foresight tokens K from 3 to 9 in Mirai-E (ablation) | Mirai-E · ImageNet 256x256, 80 epochs | FID: 5.22 -> 6.42 | Hurt |
| align layer 8 instead of layer 4 or 10 in Mirai-I (ablation) | Mirai-I · ImageNet 256x256, 80 epochs | FID: 4.98 -> 4.77 | Improved |
| align different layers between model and EMA (8 to 6) in Mirai-E (ablation) | Mirai-E · ImageNet 256x256, 80 epochs | FID: 5.22 -> 6.30 | Hurt |
| use step schedule lambda 2 to 1 instead of constant lambda (ablation) | Mirai-E · ImageNet 256x256, 300 epochs | FID: 4.96 -> 4.49 | Improved |
| use DINOv2-B instead of MAE-B as foresight encoder (ablation) | Mirai-I · ImageNet 256x256, 80 epochs | FID: 6.34 -> 4.77 | Improved |
| use larger DINOv2-L instead of DINOv2-B encoder (ablation) | Mirai-I · ImageNet 256x256, 80 epochs | FID: 4.77 -> 4.78 | No clear effect |
| restrict bidirectional encoder future access with block-causal masking (ablation) | Mirai-I · ImageNet 256x256, 50k steps | FID | Hurt |
| include the current token in aligned foresight set for Mirai-E (ablation) | Mirai-E · ImageNet 256x256, 80 epochs | FID: 5.39 -> 5.22 | Improved |
Closest papers by meaning
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.90)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.88)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.88)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 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.
