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..
Improving Diffusion-based Inverse Algorithms under Few-Step Constraint via Linear Extrapolation
Authors: Jiawei Zhang, Ziyuan Liu, Leon Yan, Gen Li, Yuantao Gu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate consistent improvements of the proposed LLE method across multiple algorithms and tasks, indicating its potential for more efficient solutions and boosted performance of diffusion-based inverse algorithms with limited steps. Codes for reproducing our experiments are available at https://github.com/weigerzan/LLE_inverse_problem. 5 Experiments In this section, we validate the effectiveness of LLE across various inverse algorithms and tasks. 5.1 Experimental setup Inverse algorithms and hyperparameters. We consider nine mainstream diffusion-based inverse algorithms with distinct formulations and design principles, including DDRM [5], DDNM [6], ΠGDM [7], DMPS [10], DPS [8], Diff PIR [9], Re Sample [12], RED-diff [11], and DAPS [14]. All these algorithms are decomposed following the canonical form described in Section 4.1, with the details provided in Appendix E. We evaluate the performance of all algorithms under total steps of S = 3, 4, 5, 7, 10, 15. Due to space constraints, we report part of the results in the main text and defer the complete results to Appendix H. For the training of LLE, we fix ω = 0.1 and utilize the schedule-free Adam W [45] to optimize the coefficients on 50 diffusion-generated samples. More training details are supplemented in Appendix G. Datasets and diffusion checkpoints. Two datasets are used: Celeb A-HQ [46] and FFHQ [47]. The checkpoint for Celeb A-HQ is obtained from [48] and we follow [6] to use the 1k test set. The checkpoint for FFHQ is obtained from [8], and the evaluation follows [12, 14] with 100 randomly selected images from the validation set. Inverse problems and metrics. We consider five inverse problems in this work: (1) gaussian deblurring (anisotropic), (2) inpainting (50% random), (3) 4 super-resolution (average pooling), (4) compressed sensing (CS) with the Walsh-Hadamard transform (50% ratio), and (5) nonlinear deblurring. Except for nonlinear deblurring, all other four tasks are linear and we apply the decoupled coefficients method proposed in Section 4.3. The primary metrics include peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [49], and Learned Perceptual Image Patch Similarity (LPIPS) [44]. Results of Fréchet Inception Distance (FID) [50] are also included in the Appendix. All of the experiments are performed on a single NVIDIA RTX 3090 GPU. |
| Researcher Affiliation | Academia | Jiawei Zhang Tsinghua University EMAIL Ziyuan Liu Tsinghua University EMAIL Leon Yan Tsinghua University EMAIL Gen Li CUHK EMAIL Yuantao Gu Tsinghua University EMAIL |
| Pseudocode | Yes | Algorithm 1 Training of LLE. Require: Pretrained diffusion model ϵθ, discrete timesteps t S, t S 1, . . . , t0, inverse algorithm in the canonical form Φti, hti, Ψti, reference samples x(1) 0 , . . . , x(N) 0 , observation function A, observation noise deviation σy. 1: for n = 1 to n = N do 2: Get observation y(n) as (5); 3: end for 4: Initialize n x(n) t S o N n=1 as x(n) t S N(0, I); 5: for i = S to i = 1 do 6: for n = 1 to n = N do 7: Calculate ˆx(n) 0,ti = hti Φti x(n) ti , A, y(n) ; 8: end for 9: Optimizing γti,0, . . . , γti,S i as (19); 10: for n = 1 to n = N do 11: Calculate x(n) 0,ti as (18); 12: Update x(n) ti 1 = Ψti( x(n) 0,ti); 13: end for 14: end for 15: return Coefficients γti,0, . . . , γti,S i for i = S, . . . , 1 Algorithm 2 Inference with LLE. Require: Pretrained diffusion model ϵθ, discrete timesteps t S, t S 1, . . . , t0, observation y, observation function A, observation noise deviation σy, inverse algorithm in the canonical form Φti, hti, Ψti, optimized coefficients γti,0, . . . , γti,S i for i = S, . . . , 1. 1: Initialize xt S N(0, I); 2: for i = S to i = 1 do 3: Calculate ˆx0,ti = hti (Φti (xti) , A, y); 4: Calculate x0,ti as (18); 5: Update xti 1 = Ψti( x0,ti); 6: end for 7: return xt0 |
| Open Source Code | Yes | Codes for reproducing our experiments are available at https://github.com/weigerzan/LLE_inverse_problem. |
| Open Datasets | Yes | Datasets and diffusion checkpoints. Two datasets are used: Celeb A-HQ [46] and FFHQ [47]. The checkpoint for Celeb A-HQ is obtained from [48] and we follow [6] to use the 1k test set. The checkpoint for FFHQ is obtained from [8], and the evaluation follows [12, 14] with 100 randomly selected images from the validation set. |
| Dataset Splits | Yes | The checkpoint for Celeb A-HQ is obtained from [48] and we follow [6] to use the 1k test set. The checkpoint for FFHQ is obtained from [8], and the evaluation follows [12, 14] with 100 randomly selected images from the validation set. |
| Hardware Specification | Yes | All of the experiments are performed on a single NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions "schedule-free Adam W [45]" as an optimizer, but does not specify software dependencies like programming languages (e.g., Python), libraries (e.g., PyTorch, TensorFlow), or their specific version numbers. |
| Experiment Setup | Yes | For the training of LLE, we fix ω = 0.1 and utilize the schedule-free Adam W [45] to optimize the coefficients on 50 diffusion-generated samples. More training details are supplemented in Appendix G. ... Appendix G.3 Details of LLE training For the training of LLE, we generate N = 50 reference samples using 999-step DDIM sampler with the corresponding diffusion model. We fix the weight of PSNR and LPIPS as ω = 0.1. The Schedule-free Adam W [45] is employed with the epochs set to 100 at each timestep ti and warmup steps set to 50. Gradients are calculated directly using the full batch. In most cases, the learning rate is set to 0.04/S, where S denotes the total number of steps used by the inverse algorithm. For Re Sample and DAPS on the nonlinear deblurring task on the FFHQ dataset, we adopt a dynamic learning rate that at timestep ti, the learning rate is set to 0.2αti+1/S. |