Research map / Image generation / Fast diffusion sampling

One Step Diffusion via Shortcut Models

ICLR 2025 · Image generation › Fast diffusion sampling · approach: Fast diffusion ODE solvers · 2 citations (OpenAlex)

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.

Introduces shortcut models, a single-network, single-run generative model conditioned on step size that enables high-quality one-step, few-step, and many-step image generation via self-consistency bootstrapping.

Problem

Diffusion and flow-matching models require many iterative denoising passes, making sampling slow; existing few-step methods rely on multi-stage distillation or fragile training schedules.

Approach

Condition a flow model on both noise level t and desired step size d, training it with a flow-matching loss at d=0 and a self-consistency loss at larger d whose targets are two consecutive half-size shortcuts, so one network learns to jump ahead in the denoising ODE end-to-end.

Key results

• On ImageNet-256 with DiT-B, shortcut models reach FID 15.5 at 128 steps, 28.3 at 4 steps, and 40.3 at 1 step, outperforming end-to-end baselines such as consistency training and live reflow.
• A DiT-XL shortcut model achieves one-step FID 10.6 and 128-step FID 3.8 on ImageNet-256.
• Shortcut model training costs roughly 16% more compute than a base diffusion model, while matching or exceeding two-stage distillation in few- and one-step settings (except progressive distillation at 1 step on some settings).

Measured effects reported in the paper

Surprising findings

• Shortcut models show slightly better FID than a comparable flow-matching model at 128 steps on CelebA-HQ and ImageNet, which the authors hypothesize is implicit regularization from the self-consistency loss.
• Flow-matching and diffusion models degrade badly at 4 and 1 steps, while the one-step shortcut model remains competitive with two-stage distillation.

Limitations

• Gap remains between many-step and one-step generation quality.
• Mapping from noise to data is entirely determined by expectation over the dataset, with no adjustable mapping as in GANs or VAEs.
• CFG scale must be specified before training.

Experimental setup

Datasets: CelebA-HQ-256 ImageNet-256 Push-T Transport

Benchmarks: CelebA-HQ-256 ImageNet-256 Push-T Transport

Metrics: FID-50k task success / performance on robotic control

Baselines: Diffusion (DiT-B) Flow Matching Reflow Progressive Distillation Consistency Distillation Consistency Training Live Reflow IBC Diffusion Policy LSTM-GMM BET

Builds on: flow matching consistency models progressive distillation DiT diffusion policy

Closest papers by meaning

• Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.91)
• Fast constrained sampling in pre-trained diffusion models (NeurIPS 2025 · similarity 0.90)
• Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.90)
• Multistep Distillation of Diffusion Models via Moment Matching (NeurIPS 2024 · similarity 0.90)
• CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.90)
• Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.90)