Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration
Authors: Yunghee Lee, Byeonghyun Pak, Junwha Hong, Hoseong Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we propose Tortoise and Hare Guidance (THG), a training-free strategy that accelerates diffusion sampling while maintaining high-fidelity generation. We demonstrate that the noise estimate and the additional guidance term exhibit markedly different sensitivity to numerical error by reformulating the classifier-free guidance (CFG) ODE as a multirate system of ODEs. Our error-bound analysis shows that the additional guidance branch is more robust to approximation, revealing substantial redundancy that conventional solvers fail to exploit. Building on this insight, THG significantly reduces the computation of the additional guidance: the noise estimate is integrated with the tortoise equation on the original, fine-grained timestep grid, while the additional guidance is integrated with the hare equation only on a coarse grid. We also introduce (i) an error-boundaware timestep sampler that adaptively selects step sizes and (ii) a guidance-scale scheduler that stabilizes large extrapolation spans. THG reduces the number of function evaluations (NFE) by up to 30% with virtually no loss in generation fidelity ( Image Reward 0.032) and outperforms state-of-the-art CFG-based training-free accelerators under identical computation budgets. Our findings highlight the potential of multirate formulations for diffusion solvers, paving the way for real-time high-quality image synthesis without any model retraining. The source code is available at https://github.com/yhlee-add/THG. (...) Using image-text pairs from the COCO 2014 dataset, we demonstrate that THG can reduce NFEs up to 30% with virtually no loss in generation fidelity ( Image Reward 0.032). THG outperforms state-of-the-art CFG-based accelerators under identical compute budgets. (...) Section 4 Experiments |
| Researcher Affiliation | Academia | Yunghee Lee Byeonghyun Pak Junwha Hong Hoseong Kim Agency for Defense Development EMAIL |
| Pseudocode | Yes | Algorithm 1 Tortoise and Hare Guidance Algorithm Algorithm 2 Look before you leap Algorithm 3 Richardson Extrapolation |
| Open Source Code | Yes | The source code is available at https://github.com/yhlee-add/THG. |
| Open Datasets | Yes | Using image-text pairs from the COCO 2014 dataset, we demonstrate that THG can reduce NFEs up to 30% with virtually no loss in generation fidelity ( Image Reward 0.032). (...) For Stable Diffusion (SD) models, we use prompt image pairs randomly sampled from COCO 2014 [27, 37]: 30,000 pairs for SD 1.5 and 1,000 pairs for SD 3.5 Large. For Audio LDM 2, we use 2,230 prompt-audio pairs from the validation set of Audio Caps [25]. |
| Dataset Splits | Yes | For Stable Diffusion (SD) models, we use prompt image pairs randomly sampled from COCO 2014 [27, 37]: 30,000 pairs for SD 1.5 and 1,000 pairs for SD 3.5 Large. For Audio LDM 2, we use 2,230 prompt-audio pairs from the validation set of Audio Caps [25]. |
| Hardware Specification | Yes | Experiments are run on a server with an AMD EPYC 74F3 26934-core CPU, 1 TB of RAM, and 8 NVIDIA A100 80GB GPUs. |
| Software Dependencies | No | We build Tortoise and Hare Guidance with Py Torch [36], Diffusers [48], and Accelerate [13]. |
| Experiment Setup | Yes | Hyperparameters (N, ω, ρ, b, ihi) are set to (50, 7.5, 1.1, 1.1, 38) for SD 1.5, (28, 3.5, 1.0, 1.2, 21) for SD 3.5 Large, and (50, 3.5, 0.9, 1.15, 39) for Audio LDM 2. |