FIG: Flow with Interpolant Guidance for Linear Inverse Problems
Authors: Yici Yan, Yichi Zhang, XIANGMING MENG, Zhizhen Zhao
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 EXPERIMENTS 4.1 EXPERIMENTAL SETUP Datasets. We conduct experiments on 3 natural image datasets: Celeb A-HQ (Karras et al., 2018), LSUN-Bedroom (Yu et al., 2015), and AFHQ-Cat (Choi et al., 2020). Metrics. For the quantitative comparison, we use the perceptual Learned Perceptual Image Patch Similarity (LPIPS) distance (Zhang et al., 2018), along with two standard metrics: peak signal-to-noise-ratio (PSNR), and structural similarity index (SSIM). Table 1: Quantitative comparison (PSNR, SSIM, LPIPS) of different algorithms for different tasks on the Celeb A-HQ 256 × 256 test dataset. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign 2Department of Statistics, University of Illinois at Urbana-Champaign 3ZJU-UIUC Institute, Zhejiang University |
| Pseudocode | Yes | Algorithm 1 Flow with Interpolant Guidance (FIG) Algorithm 2 FIG+ |
| Open Source Code | Yes | Our code is available at: https://riccizz.github.io/FIG/. |
| Open Datasets | Yes | Datasets. We conduct experiments on 3 natural image datasets: Celeb A-HQ (Karras et al., 2018), LSUN-Bedroom (Yu et al., 2015), and AFHQ-Cat (Choi et al., 2020). |
| Dataset Splits | Yes | All images are taken from the official test data splits and are preprocessed to the size of 256 × 256 × 3. ... We conduct experiments on a separate validation set (100 images from the official test data split) using different values of K and report results for K = 1 and the optimal K. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA RTX A6000 GPU for reconstructing one image. |
| Software Dependencies | No | The paper mentions using pre-trained Rectified Flow models and EDM (Karras et al., 2022) as base models, and discusses baselines like DPS, DMPS, OT-ODE, DDNM/DDNM+, and DAPS. It also references |
| Experiment Setup | Yes | For super-resolution, we apply 4× bicubic downsampling across all datasets. For deblurring tasks, we use Gaussian blurring with a kernel size of 61 × 61 and a standard deviation of 3.0, and motion blurring with the same kernel size but a standard deviation of 0.5. For inpainting, we perform random inpainting by masking out 90% of the total pixels. ... for all tasks above, we add a measurement Gaussian noise n ∼ N(0, σ2 n I) with σn = 0.05. ... Leveraging the advantages of flow matching models, we fine-tune the baseline methods to ensure they all achieve their best performance at 50 NFEs except for OT-ODE. ... The parameters c, w, K are constants, with c and K being task-specific, governing the balance between unconditional and conditional updates. |