Research map / Image generation / Fast diffusion sampling
EM Distillation for One-step Diffusion Models
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 pre-trained diffusion teacher into a one-step latent-variable generator via an EM-style maximum-likelihood procedure using joint (epsilon, z) Langevin MCMC with noise cancellation.
Problem
Diffusion models need many sampling steps; existing one-step distillation methods degrade with very few steps, need training data, or use mode-seeking divergences that miss parts of the distribution.
Approach
Frames diffusion distillation as maximum likelihood learning of a latent-variable generator via EM, where E-step samples joint (x,z) pairs initialized from the student and corrected by short-run Langevin MCMC in a reparametrized (epsilon, z) space, with accumulated Langevin noise canceled before the regression-style M-step; shows that a 1-step variant equals VSD/Diff-Instruct.
Key results
- EMD-16 reaches FID 2.20 one-step on ImageNet 64x64 class-conditional, beating the Diff-Instruct-style EMD-1 baseline (3.1) and DMD (2.62).
- EMD-16 reaches FID 6.0 one-step on ImageNet 128x128, versus 7.0 for CD and 6.3 for EMD-1.
- EMD-8 reaches FID 9.66 one-step on MSCOCO text-to-image zero-shot, better than DMD (11.49) at the same guidance setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase Langevin sampling from the one-step EMD formulation to EMD-16 (ablation) | EMD · ImageNet 64×64 class-conditional | FID: 3.1 -> 2.20 | Helped |
| Increase Langevin sampling from the one-step EMD formulation to EMD-16 (ablation) | EMD · ImageNet 128×128 class-conditional | FID: 6.3 -> 6.0 | Helped |
| Use more Langevin steps per distillation iteration (ablation) | EMD · ImageNet | FID | Helped |
| Use joint (ε,z) sampling instead of keeping z fixed (ablation) | EMD-8 · ImageNet 64×64, 100k training steps | FID: 2.829 -> 2.77 | Helped |
| Use the reparameterized (ε,z) sampler instead of sampling jointly in (x,z) (ablation) | EMD-8 · ImageNet 64×64, 100k training steps | FID: 3.11 -> 2.77 | Helped |
| Use the reparameterized (ε,z) sampler instead of sampling jointly in (x,z) (ablation) | EMD-8 · ImageNet 64×64, 100k training steps | IS: 61.08 -> 62.98 | Helped |
| Cancel all accumulated MCMC noise, including initial diffusion noise (ablation) | EMD · ImageNet | gradient variance | Helped |
| Cancel only the noise from the last MCMC step instead of all accumulated noise (ablation) | EMD · ImageNet | training stability | No clear effect |
| Increase Langevin sampling from EMD-1 to EMD-8 (ablation) | EMD · MSCOCO FID-30k text-to-image, tCFG=2 | FID: 9.78 -> 9.66 | Helped |
| Use tCFG=2 instead of tCFG=3 (ablation) | EMD-1 · MSCOCO FID-30k text-to-image | FID: 10.96 -> 9.78 | Helped |
Surprising findings
- Canceling accumulated Langevin noise markedly reduces gradient variance and speeds convergence, even though the cancellation is not provably variance-reducing.
- Joint (x,z) sampling with the x-only reparametrization underperforms sampling only epsilon, despite joint sampling being the theoretical target.
Limitations
- Requires initializing the student from the teacher; training from scratch gave poor results.
- Performance is sensitive to the choice of t* (or lambda*) for the initial denoiser.
- Extra training cost from multi-step MCMC per iteration, and MCMC step sizes need careful tuning.
Experimental setup
Datasets: ImageNet 64x64 ImageNet 128x128 LAION-Aesthetics-6.25+ MSCOCO
Benchmarks: ImageNet 64x64 ImageNet 128x128 MSCOCO-30k
Metrics: FID Inception Score Recall CLIP Score latency
Baselines: Diff-Instruct DMD CD-LPIPS BOOT UFOGen LCM-LoRA InstaFlow DPM Solver
Builds on: Diff-Instruct Variational Score Distillation EDM DMD
Closest papers by meaning
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.91)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.90)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.90)
- Efficient Dataset Distillation via Minimax Diffusion (CVPR 2024 · similarity 0.89)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · 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.