Research map / Image generation / Text-to-image diffusion
David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Diff-Instruct* (DI*), an image-data-free online RLHF post-training method for one-step text-to-image models that regularizes with a general score-based divergence, producing a 2.6B DI*-SDXL-1step model that beats 50-step FLUX-dev on Parti metrics.
Problem
Aligning one-step text-to-image generators with human preferences while preserving diversity and realism; KL-regularized RLHF (e.g., DI++) tends to mode-collapse to high-reward painting-like images.
Approach
Frame alignment as online RLHF where the regularizer is a general score-based divergence (score implicit matching style) between the one-step model and a reference diffusion model, with a tractable equivalent loss whose gradient matches the intractable objective; also incorporates classifier-free guidance as an implicit reward via explicit-implicit reward decoupling.
Key results
- DI*-SDXL-1step (2.6B) outperforms 50-step FLUX-dev (12B) on Parti in ImageReward, PickScore and CLIP score using 1.88% of inference time.
- Achieves HPSv2.1 average of 31.19 (longer training), reported as state of the art among open-source 1024px models; DDPO variant reaches 32.26.
- DI++ with KL regularization collapses to over-saturated painting-like images, while score-based DI* maintains diversity and improves preference scores.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DI* score-based preference post-training to DMD2-SDXL | DMD2-SDXL · Parti Prompts | ImageReward: 0.930 -> 1.067 | Helped |
| Apply DI* score-based preference post-training to DMD2-SDXL | DMD2-SDXL · Parti Prompts | PickScore: 0.2249 -> 0.2304 | Helped |
| Apply DI* score-based preference post-training to DMD2-SDXL | DMD2-SDXL · COCO-2017-Val 30K Prompts | COCO-FID: 15.80 -> 19.09 | Hurt |
| Apply DI* score-based preference post-training to DMD2-SDXL | DMD2-SDXL · COCO-2017-Val 30K Prompts | ImageReward: 0.822 -> 0.933 | Helped |
| Train DI*-SDXL for longer | DI*-SDXL · Parti Prompts | ImageReward: 1.067 -> 1.140 | Helped |
| Replace score-based DI* regularization with KL alignment (ablation) | DMD2-SDXL preference alignment · Parti Prompts | ImageReward: 1.067 -> 1.061 | Hurt |
| Use implicit CFG reward rather than CFG-enhanced reference diffusion (ablation) | DI*-SDXL · Parti Prompts | ImageReward: 1.067 -> 1.082 | Helped |
| Remove explicit human reward while retaining CFG reward (ablation) | DI*-SDXL · Parti Prompts | ImageReward: 1.067 -> 1.049 | Hurt |
| Remove both explicit human reward and CFG reward (ablation) | DI*-SDXL · Parti Prompts | ImageReward: 1.067 -> 0.782 | Hurt |
| Increase explicit human reward scale from 1 to 10, with CFG scale 7.5 (ablation) | DI*-SDXL · Parti Prompts | ImageReward: 1.031 -> 1.048 | Helped |
| Increase explicit human reward scale from 10 to 100, with CFG scale 7.5 (ablation) | DI*-SDXL · Parti Prompts | ImageReward: 1.048 -> 1.067 | Mixed |
| Replace SDXL reference diffusion with pre-aligned SDXL-DDPO | DI*-SDXL · Parti Prompts | ImageReward: 1.140 -> 1.210 | Helped |
Surprising findings
- KL-regularized RLHF (DI++) collapses to painting-like, oversaturated images despite high reward, whereas score-based regularization avoids this.
- Increasing explicit human reward scale degrades CLIP scores, indicating a trade-off between preference and prompt-image alignment.
Limitations
- One-step model sometimes generates inaccurate human faces and hands.
- Requires a pre-trained reward model (PickScore) rather than direct feedback data.
- Reward and CFG scales require careful tuning; large reward scale causes painting-like outputs and lower CLIP score.
Experimental setup
Datasets: LAION-Aesthetic prompts Parti prompts COCO-2014 validation prompts HPSv2 benchmark prompts
Benchmarks: Parti prompts MSCOCO-2014 30K HPSv2.1
Metrics: ImageReward Aesthetic Score PickScore CLIP score COCO-FID HPSv2.1 score inference time
Baselines: SDXL-Base SDXL-DPO SD3.5-large FLUX-dev DMD2-SDXL Diff-Instruct SIM Diff-Instruct++
Builds on: Diff-Instruct Diff-Instruct++ Score Implicit Matching DMD2 classifier-free guidance
Closest papers by meaning
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.92)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · similarity 0.91)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.90)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.90)
- GARDO: Reinforcing Diffusion Models without Reward Hacking (NeurIPS 2026 · similarity 0.90)
- Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (CVPR 2024 · similarity 0.90)
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.