Research map / Image generation / Text-to-image diffusion
VA-p: Variational Policy Alignment for Pixel-Aware Autoregressive Generation
Xinyao Liao, Qiyuan He, Kai Xu, Xiaoye Qu, Yicong Li, Wei Wei, Angela Yao
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 VA-pi post-training (pixel reconstruction reward plus prior regularization) | LlamaGen-XXL · ImageNet-1K w/o CFG | FID: 14.36 -> 7.65 | Improved |
| add VA-pi post-training | LlamaGen-XL · GenEval | Overall: 0.306 -> 0.339 | Improved |
| add VA-pi post-training | Janus-Pro 1B · GenEval | Overall: 0.725 -> 0.744 | Improved |
| use VA-pi instead of external-reward AR-GRPO | LlamaGen-XL · GenEval | Overall: 0.324 -> 0.339 | Improved |
| add VA-pi post-training on LlamaGen-XL with classifier-free guidance | LlamaGen-XL · ImageNet-1K w/ CFG | FID: 2.79 -> 2.94 | Hurt |
| add reconstruction reward without prior regularization (ablation) | VA-pi · ImageNet-1K w/o CFG | FID: 14.36 -> 38.63 | Hurt |
| add prior regularization (cross-entropy) alone without reconstruction reward (ablation) | VA-pi · ImageNet-1K w/o CFG | FID: 14.36 -> 14.17 | No clear effect |
| combine reconstruction reward with prior regularization (ablation) | VA-pi · ImageNet-1K w/o CFG | FID: 38.63 -> 7.65 | Improved |
| use contextual noise xi=0.5 instead of 0 in LlamaGen T2I post-training (ablation) | VA-pi · GenEval | Overall: 0.302 -> 0.339 | Improved |
| increase contextual noise above 0.75 (ablation) | VA-pi · GenEval | Overall: 0.339 -> 0.329 (xi=0.95) | Hurt |
| use cross-entropy instead of KL regularization with beta 0.1 (ablation) | VA-pi · ImageNet-1K w/o CFG | FID and IS | Improved |
| remove regularization (beta=0) or set it too strong (beta=1.0) (ablation) | VA-pi · ImageNet-1K w/o CFG | FID: FID 38.63 without regularization | Hurt |
Closest papers by meaning
- Group Critical-token Policy Optimization for Autoregressive Image Generation (ICLR 2026 · similarity 0.90)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- Toward Guidance-Free AR Visual Generation via Condition Contrastive Alignment (ICLR 2025 · similarity 0.88)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.88)
- Parallelized Autoregressive Visual Generation (CVPR 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.
