Research map / Image generation / Autoregressive and latent generation
NUWA-Infinity: Autoregressive over Autoregressive Generation for Infinite Visual Synthesis
Jian Liang, Chenfei Wu, Xiaowei Hu, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zicheng Liu, Yuejian Fang, Nan Duan
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 |
|---|---|---|---|
| Use autoregressive-over-autoregressive generation rather than the Taming or MaskGIT baselines | NUWA-Infinity · LHQC Text-to-Image HD | Block-FID: 24.33 -> 9.71; 38.89 -> 9.71 | Improved |
| Use autoregressive-over-autoregressive generation for images four times the training length | NUWA-Infinity · LHQC Text-to-Image HD (4096×1024) | Block-FID: 45.76 -> 15.65; 46.37 -> 15.65 | Improved |
| Use autoregressive-over-autoregressive generation instead of MaskGIT | NUWA-Infinity · LHQC Text-to-Image HD | IS: 4.61 -> 4.98 | Improved |
| Use autoregressive-over-autoregressive generation instead of MaskGIT | NUWA-Infinity · LHQC Text-to-Image HD | CLIP-SIM: 0.2754 -> 0.2807 | Improved |
| Use local autoregressive token generation instead of parallel non-autoregressive generation (ablation) | NUWA-Infinity · LHQ-V Image Animation HD | FVD: 368.16 -> 62.57 | Improved |
| Use local autoregressive token generation instead of progressive non-autoregressive generation (ablation) | NUWA-Infinity · LHQ-V Image Animation HD | FVD: 165.39 -> 62.57 | Improved |
| Add a text prompt to outpainting | NUWA-Infinity · LHQC Outpainting HD, down and up extend | Block-FID: 11.47 -> 9.84; 8.03 -> 7.43 | Improved |
| Add a text prompt to outpainting | NUWA-Infinity · LHQC Outpainting HD, right and left extend | Block-FID: 6.43 -> 6.45; 6.71 -> 6.72 | No clear effect |
| Feed positional embeddings before attention rather than after attention (ablation) | NUWA-Infinity · LHQC Text-to-Image HD | Block-FID: 10.47 -> 10.05; 18.89 -> 17.78 | Improved |
| Add cached hidden states to the Nearby Context Pool (ablation) | NUWA-Infinity · LHQC Text-to-Image HD | Block-FID: 15.80 -> 10.21; 38.32 -> 23.62 | Improved |
| Replace local autoregression with parallel non-autoregressive token generation (ablation) | NUWA-Infinity · LHQC Text-to-Image HD | Block-FID: 10.05 -> 92.34; 17.78 -> 98.67 | Hurt |
| Replace local autoregression with progressive non-autoregressive token generation (ablation) | NUWA-Infinity · LHQC Text-to-Image HD | Block-FID: 10.05 -> 19.86; 17.78 -> 38.59 | Hurt |
Later papers that compared with this method
1 later accepted paper reports NUWA-Infinity as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- MMoT: Mixture-of-Modality-Tokens Transformer for Composed Multimodal Conditional Image Synthesis (IJCV 2024 · better on 0 of 1 shared result)
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 8 of 8 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
- Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · similarity 0.91)
- InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation (NeurIPS 2025 · similarity 0.90)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis (NeurIPS 2025 · similarity 0.88)
- HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025 · similarity 0.88)
- NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis (NeurIPS 2026 · 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.
