Research map / Image generation / Fast diffusion sampling
One-step Diffusion with Distribution Matching Distillation
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.
Distills a multi-step diffusion model into a one-step image generator by matching output distributions via an approximate KL gradient (difference of real and fake diffusion scores) plus an LPIPS regression loss on precomputed noise-image pairs.
Problem
Diffusion models need dozens of network evaluations per image, making sampling too slow for interactive use; prior few-step distillation methods lag behind the teacher.
Approach
Train a one-step generator by minimizing an approximate KL divergence whose gradient is the difference between two diffusion score functions (a frozen real-data diffusion model and a jointly trained fake-data diffusion model), combined with a regression loss to teacher outputs to avoid mode collapse.
Key results
- Reaches 2.62 FID on class-conditional ImageNet 64x64 in one step, versus 6.20 for Consistency Model and 2.32 for the EDM teacher.
- Achieves 11.49 FID on zero-shot MS COCO-30k at 512x512 in one step (0.09s), versus 8.78 for SD v1.5 at 2.59s.
- Runs at 20 FPS with FP16 inference; ablation shows removing distribution matching raises ImageNet FID from 2.62 to 9.21.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add distribution matching objective (ablation) | DMD · CIFAR-10 | FID: 3.82 -> 2.66 | Helped |
| Add distribution matching objective (ablation) | DMD · ImageNet-64×64 | FID: 9.21 -> 2.62 | Helped |
| Add regression loss (ablation) | DMD · CIFAR-10 | FID: 5.58 -> 2.66 | Helped |
| Add regression loss (ablation) | DMD · ImageNet-64×64 | FID: 5.61 -> 2.62 | Helped |
| Use Eq. 8 sample weighting instead of σt/αt weighting (ablation) | DMD · CIFAR-10 | FID: 3.60 -> 2.66 | Helped |
| Use Eq. 8 sample weighting instead of σ³t/αt weighting (ablation) | DMD · CIFAR-10 | FID: 3.71 -> 2.66 | Helped |
| Distill EDM into a one-step generator | EDM · ImageNet-64×64 | FID: 2.32 -> 2.62 | Hurt |
| Distill Stable Diffusion v1.5 into a one-step generator at guidance scale 3 | Stable Diffusion v1.5 · MS COCO 2014-30k, zero-shot | FID: 8.78 -> 11.49 | Hurt |
| Distill Stable Diffusion v1.5 into a one-step generator at guidance scale 8 | Stable Diffusion v1.5 · MS COCO-30K | FID: 13.45 -> 14.93 | Hurt |
Surprising findings
- Pre-computed teacher noise-image pairs with a simple regression loss suffice for distillation once distribution matching is added, even when the pairs cost under 1% of training compute on CIFAR-10.
Limitations
- Slight quality gap relative to multi-step teacher sampling with 100 or 1000 steps.
- Fine-tuning both the fake score network and generator causes significant training memory usage.
- Regression loss requires precomputing teacher noise-image pairs, which adds cost.
Experimental setup
Datasets: ImageNet CIFAR-10 LAION-Aesthetics-6.25+ LAION-Aesthetics-6+ MS COCO 2014
Benchmarks: ImageNet 64x64 CIFAR-10 MS COCO 2014-30k zero-shot
Metrics: FID CLIP Score latency
Baselines: Consistency Model TRACT Progressive Distillation InstaFlow UFOGen LCM-LoRA DPM++ UniPC BigGAN-deep GigaGAN StyleGAN-T
Builds on: Variational Score Distillation Score-SDE EDM Consistency Models pix2pix
Closest papers by meaning
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.93)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.92)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.92)
- Multistep Distillation of Diffusion Models via Moment Matching (NeurIPS 2024 · similarity 0.92)
- Improved Distribution Matching Distillation for Fast Image Synthesis (NeurIPS 2024 · similarity 0.92)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.91)
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.