Research map / Image generation / Autoregressive and latent generation
G3AFT: Glance Guided Gradient Aligned Fine-Tuning for Visual Autoregressive Models
Jiayi Zhang, Xiefan Guo, Xinzhu Ma, Di Huang
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 |
|---|---|---|---|
| end-to-end fine-tuning with GSGR+NSGP versus base model | Janus-Pro-1B · HPSv2 Average | HPSv2 reward: 26.30 -> 28.03 | Improved |
| G3AFT versus GRPO | Janus-Pro-1B · HPSv2 Average | HPSv2 reward: 27.70 -> 28.03 | Improved |
| G3AFT versus GRPO on Photo | Janus-Pro-1B · HPSv2 Photo | HPSv2 reward: 27.42 -> 26.64 | Hurt |
| aesthetic reward fine-tuning | Janus-Pro-1B · Animals | aesthetic score: 5.37 -> 7.05 | Improved |
| add GSGR periodic gradient recording instead of later-step tokens (ablation) | G3AFT · HPSv2 Average | HPSv2 reward: 27.16 -> 27.43 | Improved |
| add GSGR (memory) (ablation) | G3AFT · HPSv2 | peak GPU memory: 67.05 GB -> 53.90 GB | Improved |
| add GSGR on Photo category (ablation) | G3AFT · HPSv2 Photo | HPSv2 reward: 26.29 -> 26.07 | Hurt |
| add NSGP to pure softmax relaxation (ablation) | G3AFT · HPSv2 Average | HPSv2 reward: 27.84 -> 28.03 | Improved |
| use pure softmax instead of Gumbel-softmax (no NSGP) (ablation) | G3AFT · HPSv2 Average | HPSv2 reward: 27.72 -> 27.84 | Improved |
| entropy-driven selection at 33.3% ratio instead of strided at 50% (ablation) | G3AFT · HPSv2 Average | HPSv2 reward: 28.03 -> 28.05 | No clear effect |
Closest papers by meaning
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- ScalingAR: Scaling Confidence for Autoregressive Image Generation (ICML 2026 · similarity 0.89)
- Next Patch Prediction for AutoRegressive Visual Generation (AAAI 2026 · similarity 0.88)
- Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling (ECCV 2026 · similarity 0.88)
- MoSs: Mixture of Scales for Efficient High-Resolution Autoregressive Image Generation (AAAI 2026 · similarity 0.88)
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · 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.
