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..

Learning to Integrate Diffusion ODEs by Averaging the Derivatives

Authors: Wenze Liu, Xiangyu Yue

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on CIFAR-10 [37] and Image Net-256 256 [38] datasets, using EDM [6] and Si T [39] as teacher diffusion models, respectively. Our experiments demonstrate that diffusion models can be efficiently converted to their secant version, with significantly slower accuracy degradation compared to conventional numerical solvers as the step number decreases. On CIFAR-10, our approach achieves FID scores of 2.14 with 10 steps. For Image Net-256 256 in latent space, we obtain a 4-step FID of 2.78 and an 8-step FID of 2.33. With the guidance interval technique [40], the performance is further improved to 2.27 with 4 steps and 1.96 with 8 steps.
Researcher Affiliation Academia Wenze Liu Xiangyu Yue MMLab, The Chinese University of Hong Kong
Pseudocode Yes Algorithm 1 Secant Distillation by Estimating the Interior Point (SDEI) Algorithm 2 Secant Training by Estimating the End Point (STEE) Algorithm 3 Secant Training by Estimating the Interior Point (STEI) Algorithm 4 Secant Distillation by Estimating the End Point (SDEE) Algorithm 5 Sampling of t, s in SDEI and STEI for EDM Algorithm 6 Sampling of t, s in SDEE and STEE for EDM Algorithm 7 Sampling of t, s in SDEI and STEI for Di T Algorithm 8 Sampling of t, s in SDEE and STEE for Di T
Open Source Code Yes Code is available at https://github.com/poppuppy/secant-expectation.
Open Datasets Yes We evaluate our method on CIFAR-10 [37] and Image Net-256 256 [38] datasets, using EDM [6] and Si T [39] as teacher diffusion models, respectively.
Dataset Splits Yes We evaluate our method on CIFAR-10 [37] and Image Net-256 256 [38] datasets, using EDM [6] and Si T [39] as teacher diffusion models, respectively.
Hardware Specification Yes All the experiments can be done on a sever with 8 NVIDIA A100 GPUs.
Software Dependencies No The paper does not explicitly state specific software dependencies with version numbers. It mentions the use of 'Py Torch' implicitly but without a version.
Experiment Setup Yes The detailed experimental configurations are presented in Table 9.