Research map / Image generation / Fast diffusion sampling
Consistency Models Made Easy
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 Easy Consistency Tuning (ECT), which fine-tunes a pretrained diffusion model into a consistency model via a continuous-time differential consistency condition, achieving strong 1- and 2-step FIDs at a fraction of prior training cost.
Problem
Consistency models offer fast few-step sampling but are very expensive and finicky to train from scratch or via distillation; the training objective becomes unstable as the discretization interval shrinks.
Approach
Express consistency training in continuous time via a differential consistency condition, show diffusion training is the loose-discretization special case (r=0), and progressively tighten the condition by shrinking the gap between t and r, starting from a pretrained diffusion model without a frozen teacher.
Key results
- ECM achieves 2-step FID of 2.73 on CIFAR-10 in one hour on a single A100 GPU, beating CD's 2.93 (800k iters at batch 512).
- On ImageNet 64x64 with a 102.4M-image budget, ECM-XL reaches 1-step FID 2.49 and 2-step FID 1.67, versus iCT-deep 2.77 (2-step).
- ECT uses about one third of iCT's total training FLOPs while beating iCT in 2-step FID; the tuning stage is 0.60% to 1.91% of pretraining budget.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ECT instead of CD (LPIPS) | CIFAR-10 2-step generation · CIFAR-10 | FID: 2.93 vs 2.20 | Helped |
| Use ECT instead of iCT-deep | CIFAR-10 2-step generation · CIFAR-10 | FID: 2.24 -> 2.11 | Helped |
| Increase ECT training from 100k to 400k iterations | CIFAR-10 1-step generation · CIFAR-10 | FID: 4.54 -> 3.60 | Helped |
| Increase model size from ECM-S to ECM-XL | ImageNet 64×64 2-step generation · ImageNet 64×64 | FID: 3.18 -> 1.96 | Helped |
| Add continuous-time schedule (ablation) | ECT fine-tuning from EDM2 · ImageNet 64×64 | FID (1-step): 21.09 -> 14.34 | Helped |
| Add continuous-time schedule (ablation) | ECT fine-tuning from EDM2 · ImageNet 64×64 | FID (2-step): 4.39 -> 4.33 | Helped |
| Increase dropout to 0.40 (ablation) | ECT fine-tuning from EDM2 · ImageNet 64×64 | FID (1-step): 14.34 -> 9.28 | Helped |
| Increase dropout to 0.40 (ablation) | ECT fine-tuning from EDM2 · ImageNet 64×64 | FID (2-step): 4.33 -> 3.22 | Helped |
| Use timestep weighting 1/t² + 1/σ_data² (ablation) | ECT fine-tuning from EDM2 · ImageNet 64×64 | FID (1-step): 9.28 -> 5.51 | Helped |
| Use timestep weighting 1/t² + 1/σ_data² (ablation) | ECT fine-tuning from EDM2 · ImageNet 64×64 | FID (2-step): 3.22 -> 3.18 | Helped |
| Use sigmoid mapping n(t) instead of constant n(t) ≡ 1 (ablation) | ECT · CIFAR-10 | FID (1-step): 4.06 -> 3.86 | Helped |
| Start directly with Δt ≈ dt (q=256) instead of gradually shrinking Δt (ablation) | ECT · CIFAR-10 | FID (1-step): 3.60 -> 3.86 | Hurt |
Surprising findings
- 2-step ECM models from smaller size can outperform 1-step larger models (e.g., 498M ECM-M 2-step vs 1.1B ECM-XL 1-step).
- Adaptive weighting with p=1 on flow matching pretraining already yields strong few-step sampling.
Limitations
- Requires a training dataset for tuning, unlike data-free distillation methods.
- Scaling-law analysis limited to CIFAR-10 and ImageNet 64x64 with a few model sizes.
- Pretraining-tuning design depends on hyperparameters (q, d, dropout) tuned per dataset.
Experimental setup
Datasets: CIFAR-10 ImageNet 64x64
Benchmarks: CIFAR-10 ImageNet 64x64
Metrics: FID FD_DINOv2 NFE
Baselines: iCT iCT-deep Consistency Distillation (CD) Score SDE EDM EDM (DPM-Solver-v3) TRACT Diff-Instruct GET PD BOOT DFNO
Builds on: consistency models improved Consistency Training (iCT) EDM EDM2 Consistency Distillation
Closest papers by meaning
- Simplifying, Stabilizing and Scaling Continuous-time Consistency Models (ICLR 2025 · similarity 0.91)
- Improving Consistency Models with Generator-Augmented Flows (ICML 2025 · similarity 0.91)
- Truncated Consistency Models (ICLR 2025 · similarity 0.90)
- Convergence of Consistency Model with Multistep Sampling under General Data Assumptions (ICML 2025 · similarity 0.89)
- Sampling is as easy as keeping the consistency: convergence guarantee for Consistency Models (ICML 2024 · similarity 0.88)
- Improved Discretization Complexity Analysis of Consistency Models: Variance Exploding Forward Process and Decay Discretization Scheme (ICML 2025 · similarity 0.88)
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.