Research map / Image generation / Text-to-image diffusion
AdaNAT: Exploring Adaptive Policy for Token-Based Image Generation
Zanlin Ni, Yulin Wang, Renping Zhou, Rui Lu, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Yuan Yao, Gao 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 |
|---|---|---|---|
| make the generation policy learnable (policy network instead of manual schedules) (ablation) | AdaNAT-S (T=4) · ImageNet 256x256 | FID-50K: 7.65 -> 5.40 | Improved |
| make the learned policy adaptive to each sample (ablation) | AdaNAT-S (T=4) · ImageNet 256x256 | FID-50K: 5.40 -> 4.54 | Improved |
| use FID as the reward with an adaptive policy (ablation) | AdaNAT-L (T=8) · ImageNet 256x256 | FID-50K: 55.4 (Fail) | Hurt |
| use FID as the reward with a non-adaptive policy (ablation) | AdaNAT-L (T=8) · ImageNet 256x256 | FID-50K vs visual quality: 2.86 -> 2.56 FID (adversarial vs FID-reward) | Mixed |
| use a pre-trained image reward model as reward (ablation) | AdaNAT · ImageNet 256x256 | sample diversity | Hurt |
| use adversarial discriminator reward instead of FID or pretrained reward model (ablation) | AdaNAT · ImageNet 256x256 | fidelity/diversity balance | Improved |
| apply AdaNAT policy to Muse on CC3M | Muse · CC3M | FID-30K: 7.67 -> 6.83 | Improved |
| use adversarial reward instead of FID reward vs AutoNAT (ablation) | AdaNAT-L · ImageNet 256x256 (T=8) | FID-50K: 2.56 -> 2.86 (FID-based vs Adv-based); AutoNAT-L 2.68 | Hurt |
Closest papers by meaning
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- Policy Optimized Text-to-Image Pipeline Design (NeurIPS 2025 · similarity 0.87)
- Generation Navigator: A State-Aware Agentic Framework for Image Generation (NeurIPS 2026 · similarity 0.87)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.86)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.86)
- Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models (NeurIPS 2026 · 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.
