Research map / Image generation / Autoregressive and latent generation
Attribute Token Arithmetic: Disentangled and Continuous Semantic Control for Visual Autoregressive Models
Xindi Yang, Yicheng Wu, Cheng Zhang, Jianfei Cai, Tien-Tsin Wong
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 |
|---|---|---|---|
| replace raw text-token difference with learned disentangled directions (ablation) | ATA (Infinity-2B) · GEdit | Semantic Consistency: 5.048 -> 5.374 | Improved |
| replace raw text-token difference with learned disentangled directions (ablation) | ATA (Infinity-2B) · GEdit | Perceptual Quality: 7.447 -> 7.584 | Improved |
| replace raw text-token difference with learned disentangled directions (ablation) | ATA (Infinity-2B) · GEdit | Disentanglement: 5.296 -> 5.455 | Improved |
| use ATA token directions instead of Concept Slider LoRA on same backbone (ablation) | Infinity-2B · Control Dataset | Mean DeltaVQA: 0.532 -> 0.567 | Improved |
| use ATA token directions instead of Concept Slider LoRA on same backbone (ablation) | Infinity-2B · Control Dataset | Mean I-LPIPS: 0.879 -> 0.895 | Improved |
| perturb prompt wording at test time (ablation) | ATA (Infinity-2B) · Control Dataset | Mean DeltaVQA: 0.567 -> 0.541 | Hurt |
| perturb prompt wording at test time (ablation) | ATA (Infinity-2B) · Control Dataset | Mean I-LPIPS: 0.895 -> 0.864 | Hurt |
| apply attribute direction to identity token instead of attribute token (ablation) | ATA · qualitative | identity preservation | Hurt |
| apply direction to unrelated token such as ground (ablation) | ATA · qualitative | edit effect | Hurt |
| use ATA instead of VAREdit at extreme intensity | ATA · Control Dataset | DeltaVQA: +6.95% -> +14.58% | Improved |
| use ATA instead of Qwen-Image at moderate intensity | ATA · GEdit | Semantic Consistency: +6.20% -> +16.33% | Improved |
| use ATA instead of VAREdit at moderate intensity | ATA · GEdit | Perceptual Quality: +5.75% -> +3.16% | Hurt |
Closest papers by meaning
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.88)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.88)
- Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions (CVPR 2025 · similarity 0.87)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.87)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.87)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 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.
