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..

Latent Refinement via Flow Matching for Training-free Linear Inverse Problem Solving

Authors: Hossein Askari, Yadan Luo, Hongfu Sun, Fred Roosta

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our proposed method outperforms state-of-the-art latent diffusion solvers in reconstruction quality across most tasks. The code will be publicly available at Git Hub. Our empirical evaluations demonstrate that images inferred via latent ODE sampling along conditional OT paths exhibit superior perceptual quality compared to those generated through latent diffusion-based probability paths. Empirical: We validate the performance of LFlow through extensive experiments on image reconstruction tasks, including deblurring, super-resolution, and inpainting, achieving state-of-the-art results without requiring substantial problem-specific hyperparameter tuning.
Researcher Affiliation Academia 1The University of Queensland 2ARC Training Centre for Information Resilience (CIRES) EMAIL
Pseudocode Yes Algorithm 1 LFlow Sampling: Posterior-Guided Latent ODE Inference for Linear Inverse Problems
Open Source Code No The code will be publicly available at Git Hub.
Open Datasets Yes Datasets and Tasks We evaluate our method on three datasets: FFHQ [69], Image Net [70], and Celeb A-HQ [71], each containing images with a resolution of 256  256  3 pixels.
Dataset Splits Yes We use 200 randomly selected validation samples per dataset. All images are normalized to the range [ 1, 1]. We present our findings on several linear inverse problem tasks, including Gaussian deblurring, motion deblurring, super-resolution, and box inpainting. The measurement operators are configured as follows: (i) Gaussian deblurring convolves images with a Gaussian blurring kernel of size 61  61 and a standard deviation of 3.0 [29]; (ii) Motion Deblurring uses motion blur kernels that are randomly generated with a size of 61  61 and an intensity value of 0.5 [29]; (iii) Super-Resolution (SR) involves bicubic downsampling by a factor of 4; and (iv) Box Inpainting simulates missing data by masking a 128  128 pixel box, which is randomly positioned around the center with a margin of [16, 16] pixels, following the methodology of [21].
Hardware Specification Yes All experiments were conducted on a single NVIDIA 3090 GPU with a batch size of 1.
Software Dependencies No After evaluating various numerical ODE integration methods, we selected adaptive_heun as the default solver, utilizing its reliable implementation from the open-source torchdiffeq library [62].
Experiment Setup Yes For the solver parameters, we set the absolute and relative tolerances (atol and rtol) to 103 for inpainting and motion deblurring tasks, and to 105 for Gaussian deblurring and super-resolution tasks. We set the hyperparameters to K = 2 and ts = 0.8 for all tasks.