Research map / Reinforcement learning / Imitation learning
FAIL: Flow Matching Adversarial Imitation Learning for Image Generation
Yeyao Ma, Chen Li, Xiaosong Zhang, Han Hu, Weidi Xie
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 |
|---|---|---|---|
| use FAIL-PG instead of SFT on expert data | FLUX.1-dev · UniGen-Bench (512px) | score: 58.22 -> 64.75 | Improved |
| train RLHF-PD for 400 instead of 50 steps (ablation) | RLHF-PD · UniGen-Bench | score: 59.76 -> 57.97 | Hurt |
| replace Online DPO with FAIL-PG (learned discriminator) | FLUX.1-dev · UniGen-Bench | score: 62.83 -> 64.75 | Improved |
| extend FAIL-PG training beyond 450 steps (ablation) | FAIL-PG · UnifiedReward over training | reward | Hurt |
| extend FAIL-PD training to 2,000 steps (ablation) | FAIL-PD · UnifiedReward over training | reward: 3.4759 at step 2,000 | Improved |
| use DINOv3 visual foundation model discriminator instead of Qwen3-VL (ablation) | FAIL-PD · UniGen-Bench | score: 62.17 -> 60.65 | Hurt |
| use FLUX.1 backbone discriminator instead of Qwen3-VL (ablation) | FAIL-PD · HPSv3 | score: 10.82 -> 12.35 | Improved |
| vary discriminator type (DINOv3, FLUX.1, Qwen3-VL) (ablation) | FAIL-PG · UniGen-Bench | score: 64.23, 63.33, 64.75 | No clear effect |
| add FAIL-PD loss to Reward Gradient optimization | Reward Gradient (HPSv3) · UniGen-Bench / UnifiedReward | score: 58.50 -> 60.88 (UniGen); 3.4254 -> 3.4832 (URScore) | Improved |
| add FAIL-PG score to FPO | FPO (HPSv3 reward) · UniGen-Bench | score: 61.27 -> 64.27 | Improved |
| apply FAIL-PG to Xomni instead of SFT | Xomni (discrete AR) · UniGen-Bench | score: 41.36 -> 54.24 | Improved |
| apply FAIL-PG instead of SFT to Wan2.1-1.3B | Wan2.1-1.3B · VBench Semantic Score | score: 46.27 -> 51.75 | Improved |
Closest papers by meaning
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.88)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.87)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.87)
- Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models (NeurIPS 2026 · similarity 0.87)
- Diffusing States and Matching Scores: A New Framework for Imitation Learning (ICLR 2025 · similarity 0.87)
- Exploring Gradient Explosion in Generative Adversarial Imitation Learning: A Probabilistic Perspective (AAAI 2024 · 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.
