Research map / Image generation / Text-to-image diffusion
Flash-DMD: Towards High-Fidelity Few-Step Image Generation with Efficient Distillation and Joint Reinforcement Learning
Guanjie Chen, Shirui Huang, Yifu Sun, Kai Liu, Jianchen Zhu, Xiaoye Qu, Yu Cheng, Peng Chen
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 Pixel-GAN to DMD2 (ablation) | DMD2 SDXL · COCO-10k | ImageReward: 0.8748 -> 0.8918 | Improved |
| use aggressive TTUR (lower score-estimator update ratio) in DMD2 (ablation) | DMD2 SDXL · COCO-10k | ImageReward: 0.8748 -> 0.8871 | Improved |
| add timestep-aware loss decoupling on top of Pixel-GAN and aggressive TTUR (ablation) | Flash-DMD SDXL · COCO-10k | ImageReward: 0.8748 -> 0.9509 (24k -> 1k steps) | Improved |
| add joint reinforcement learning with distillation (ablation) | Flash-DMD SDXL · COCO-10k | ImageReward: 0.9509 -> 1.0035 | Improved |
| increase training steps from 1k to 8k (ablation) | Flash-DMD SDXL · COCO-10k | ImageReward: 0.9509 -> 0.9740 | Improved |
| use TTUR 5 instead of 1-2 in Flash-DMD (ablation) | Flash-DMD SDXL · COCO-10k | ImageReward: 0.9426 (TTUR5-18k) vs 0.9509 (TTUR1-1k) | No clear effect |
| use RL sampling over all noise timesteps instead of only high-noise ones (ablation) | Flash-DMD phase 2 · COCO-10k | ImageReward: 1.0004 -> 0.9421 | Hurt |
| apply RL as separate post-training instead of jointly with distillation (ablation) | Flash-DMD phase 2 · COCO-10k | HPSv2: 0.2931 -> 0.2882 | Hurt |
| alternate RL and DM updates at 5:1 ratio rather than 1:1 (ablation) | Flash-DMD phase 2 · COCO-10k | ImageReward: 0.9135 -> 0.9808 | Improved |
| raise RL:DM ratio further to 10:1 (ablation) | Flash-DMD phase 2 · COCO-10k | HPSv2: 0.2904 (5:1) -> 0.2874 (10:1) | Hurt |
| add Pixel-GAN loss during RL stage (ablation) | Flash-DMD phase 2 · COCO-10k | ImageReward: 0.9421 -> 0.9678 | Improved |
| use Flash-DMD instead of DMD2 at TTUR=2 (ablation) | SDXL distillation · COCO-10k | training stability | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Hyper-SDXL — Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis (NeurIPS 2024) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.90)
- Improved Distribution Matching Distillation for Fast Image Synthesis (NeurIPS 2024 · similarity 0.90)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.90)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.89)
- D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models (NeurIPS 2026 · similarity 0.89)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.89)
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.
