Research map / Image generation / Autoregressive and latent generation
Unlocking Any-Order Generation in Pretrained Autoregressive Image Models
Rishav Pramanik, Marco Pedersoli, Zhaozheng Yin
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a method to unlock arbitrary-order (any-order) generation in already-pretrained raster-scan autoregressive image models, enabling inpainting, outpainting, and editing at inference time.
Problem
Raster-scan autoregressive image generators are fixed to a single factorization of the image distribution, so they cannot inpaint, outpaint, or edit at inference; existing any-order solutions (including discrete diffusion) train from scratch on mixed token orders at roughly 3x compute.
Approach
Adapts existing pretrained raster-scan AR image models to any-order generation rather than training from scratch on mixed orderings, avoiding the ~3x training compute cost of prior any-order approaches. Details are not available in the provided abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt pretrained raster-scan checkpoint rather than train random init at matched total epochs (ablation) | any-order recipe on LlamaGen-L · ImageNet class-conditional | FID: 5.60 -> 4.24 | Improved |
| adapt pretrained raster-scan checkpoint rather than train random init at matched total epochs (ablation) | any-order recipe on LlamaGen-XL · ImageNet class-conditional | FID: 4.81 -> 3.91 | Improved |
| add ordering curriculum (linear alpha ramp) instead of alpha = 1 from step 0 (ablation) | any-order adaptation of LlamaGen-L · ImageNet class-conditional | FID: 5.86 -> 4.24 | Improved |
| add future-aware positional embeddings (ablation) | any-order adaptation · training loss | training loss convergence | Improved |
| train embedding table at 10x base learning rate (ablation) | future-aware embedding table · any-order adaptation | convergence and any-order quality | Improved |
| add KL regularization toward frozen raster-scan teacher (ablation) | any-order adaptation of LlamaGen-L · ImageNet class-conditional | FID: 4.24 -> 9.62 | Hurt |
| inject future-aware embedding at every layer (AdaLN-style) instead of input layer only (ablation) | future-aware embedding · design choice | quality | No clear effect |
| adapt pretrained raster model to any-order | LlamaGen-XL · ImageNet 256 class-conditional | FID: 2.63 -> 3.91 | Hurt |
| adapt pretrained raster model to any-order | RAR-L · ImageNet 256 class-conditional | FID: 1.70 -> 3.63 | Hurt |
| use any-order adapted model instead of MaskGIT (cosine) for inpainting | ImageNet inpainting · ImageNet inpainting | FID: 10.44 -> 6.21 | Improved |
| use Halton schedule for MaskGIT instead of cosine | MaskGIT baseline · ImageNet inpainting | FID: 10.44 -> 7.82 | Improved |
| adapted model versus MAR-L on inpainting and outpainting | ImageNet inpainting/outpainting · ImageNet | FID: 6.63 -> 6.21 inpainting; 6.86 -> 5.84 outpainting | No clear effect |
Limitations
- Only the abstract is available; method details and results are unknown.
Experimental setup
Baselines: discrete diffusion models
Closest papers by meaning
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.89)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.88)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.88)
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.
