Research map / Image generation / Fast diffusion sampling

Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 0 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.

Proposes Dual-Rate Diffusion, which interleaves a heavy context encoder evaluated sparsely with a light denoising model run at every step, and extends it to Moment Matching Distillation, reporting comparable ImageNet FID at 2-4x lower FLOPs.

Problem

Diffusion model sampling requires repeated evaluation of a heavy network at every step, making inference expensive even for few-step distilled models.

Approach

Split the denoiser into a heavy context encoder evaluated only every K-th step and a light denoising model that reuses the encoder's multi-level features at every step, trained jointly with a Markovian objective (and rollout-based training for distillation).

Key results

• Dual-Rate Diffusion (K=16, 512 denoising steps) reaches FID 1.12 on ImageNet 64x64 vs 1.25 for the baseline (guidance w=1) at 13.14 vs 55.5 TFLOPs.
• Dual-Rate MMD with 8 NFE reaches FID 1.17 on ImageNet 64x64 vs 1.24 for standard MMD at 8 steps (0.79 vs 0.87 TFLOPs in table), and 1.90 vs 1.49 at 128x128 (worse than standard MMD at 8 steps).
• Dual-Rate Diffusion at 128x128 with 256 NFE reaches FID 1.89 at 13.98 TFLOPs vs baseline FID 1.50 at 70.37 TFLOPs.

Surprising findings

• Dual-Rate Diffusion on ImageNet 64x64 outperforms the baseline FID with a roughly 4x smaller FLOP budget.
• Applying classifier-free guidance to the teacher during MMD training did not improve student performance and sometimes caused instability.

Limitations

• Increases training cost per iteration because both context encoder and denoiser are evaluated each step.
• Requires Markovian stochastic processes and student rollouts in distillation to avoid distribution mismatch.
• Results limited to ImageNet at 64x64 and 128x128 resolutions with a UViT architecture.

Experimental setup

Datasets: ImageNet 64x64 ImageNet 128x128

Benchmarks: ImageNet 64x64 ImageNet 128x128

Metrics: FID NFE TFLOPs

Baselines: VDM++ EDM2-XL SiD2 CTM sCD-XXL Moment Matching DMD2

Builds on: Moment Matching Distillation self-conditioning EDM2 UViT classifier-free guidance

Closest papers by meaning

• Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.92)
• Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models (ICML 2025 · similarity 0.90)
• A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
• ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.90)
• Neural Diffusion Models (ICML 2024 · similarity 0.90)
• Diffusion Models Without Attention (CVPR 2024 · similarity 0.90)