Research map / Image generation / Fast diffusion sampling
AdaGen: Learning Adaptive Policy for Image Synthesis
Zanlin Ni, Yulin Wang, Yue Hua, Renping Zhou, Jiayi Guo, Jun Song, Bo Zheng, Gao Ming 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 |
|---|---|---|---|
| learn the generation schedule instead of using the hand-designed one (ablation) | MaskGIT-S · ImageNet 256 (T=4) | FID-50K: 7.65 -> 5.40 | Improved |
| learn the generation schedule instead of using the hand-designed one (ablation) | DiT-XL · ImageNet 256 (T=4) | FID-50K: 9.71 -> 6.03 | Improved |
| make the learned policy sample-adaptive (ablation) | MaskGIT-S · ImageNet 256 (T=4) | FID-50K: 5.40 -> 4.54 | Improved |
| make the learned policy sample-adaptive (ablation) | DiT-XL · ImageNet 256 (T=4) | FID-50K: 6.03 -> 5.31 | Improved |
| add repeated sampling selected by the adversarial reward model (ablation) | AdaGen-MaskGIT-L · ImageNet 256 (T=32) | FID-50K: 2.28 -> 2.07 | Improved |
| add lookahead sampling with the value network (K=2) to repeated sampling (ablation) | AdaGen-MaskGIT-L · ImageNet 256 (T=32) | FID-50K: 2.07 -> 1.94 | Improved |
| add repeated sampling selected by the adversarial reward model (ablation) | AdaGen-DiT-XL · ImageNet 256 (T=16) | FID-50K: 2.19 -> 2.06 | Improved |
| increase repeated sampling trials from 0 to 3 (with lookahead) (ablation) | AdaGen-MaskGIT-L · ImageNet 256 (T=32) | FID-50K: 2.28 -> 1.81 | Improved |
| increase repeated sampling trials beyond 3 to 6 (ablation) | AdaGen-MaskGIT-L · ImageNet 256 (T=32) | FID-50K: 1.81 -> 1.90 | Hurt |
| increase repeated sampling trials from 0 to 5 (ablation) | AdaGen-DiT-XL · ImageNet 256 (T=16) | FID-50K: 2.19 -> 1.91 | Improved |
| use FID as the reward for an adaptive policy (ablation) | AdaGen-MaskGIT-L · ImageNet 256 (T=8) | FID-50K: 2.56 -> 55.4 (Fail) | Hurt |
| reduce inference steps (gain of AdaGen grows as steps drop) | AdaGen-DiT-XL · ImageNet 256 | FID-50K: -1.12 at 16 steps, -2.36 at 8 steps, -4.40 at 4 steps | Improved |
Closest papers by meaning
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 2025 · similarity 0.89)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.86)
- Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (NeurIPS 2024 · similarity 0.86)
- Adaptive Divergence Regularized Policy Optimization for Fine-tuning Generative Models (NeurIPS 2025 · similarity 0.86)
- Stochastic Interpolants via Conditional Dependent Coupling (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.
