Research map / Image generation / Autoregressive and latent generation
MAGE: MAsked Generative Encoder To Unify Representation Learning and Image Synthesis
Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, Dilip Krishnan
This paper tried a common change: Add a contrastive loss
Add contrastive loss to reconstructive loss · ImageNet-1K linear probing · top-1 accuracy: 73.3 -> 77.1 Improved
What 1,317 accepted papers report for this change
- 1,253 improved
- 22 no clear effect
- 16 mixed
- 26 hurt
- Beyond Looking Up, Try Looking Around: Harmonizing Global Structure and Local Consistency in Optimal Transport for Short Text Clustering ICML 2026 · add contrastive learning · SearchSnippets · ACC: 56.65 -> 80.12Improved
- SleepLM: Natural-Language Intelligence for Human Sleep ICML 2026 · add contrastive objective to captioning (CoCa vs Cap) · sleep event IoU · IoU: 29.2 -> 27.4Hurt
- LIMSSR: LLM-Driven Sequence-to-Score Reasoning under Training-Time Incomplete Multimodal Observations ICML 2026 · add contrastive loss L_con · FS1000 average · Spearman correlation: 0.759 -> 0.789Improved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
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 variable rather than fixed masking ratios (ablation) | MAGE · ImageNet-1K class-unconditional generation | FID: N/A -> 12.5 | Improved |
| Use variable rather than fixed masking ratios (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 71.5 -> 72.2 | Improved |
| Use quantized tokens instead of unquantized features (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 49.5 -> 74.7 | Improved |
| Add image-specific [C] padding instead of universal [MASK] padding (ablation) | MAGE · ImageNet-1K class-unconditional generation | FID: 12.4 -> 11.6 | Improved |
| Add image-specific [C] padding instead of universal [MASK] padding (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 72.5 -> 73.3 | Improved |
| Use weak rather than strong augmentation (ablation) | MAGE · ImageNet-1K class-unconditional generation | FID: 11.1 -> 8.67 | Improved |
| Use weak rather than strong augmentation (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 74.7 -> 70.5 | Hurt |
| Train for more pre-training epochs (ablation) | MAGE · ImageNet-1K class-unconditional generation | FID: 12.2 -> 11.6 -> 11.1 | Improved |
| Train for more pre-training epochs (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 72.2 -> 73.3 -> 74.7 | Improved |
| Add contrastive loss to reconstructive loss (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 73.3 -> 77.1 | Improved |
| Remove reconstructive loss, leaving contrastive loss only (ablation) | MAGE · ImageNet-1K linear probing | top-1 accuracy: 77.1 -> 72.9 | Hurt |
| Reduce maximum masking ratio (ablation) | MAGE-C · ImageNet-1K class-unconditional generation | FID: 14.1 -> 23.5 -> 27.0 | Hurt |
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
- Fixed-Point Masked Generative Modeling (NeurIPS 2026 · similarity 0.87)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.87)
- BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities (ICLR 2025 · similarity 0.86)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.86)
- Plug-and-Play Context Feature Reuse for Efficient Masked Generation (NeurIPS 2025 · similarity 0.86)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · 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.
