PSC: Posterior Sampling-Based Compression
Authors: Noam Elata, Tomer Michaeli, Michael Elad
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We begin with an evaluation of PSC on 256 256 color images from the Image Net (Deng et al., 2009) dataset... We compare distortion (PSNR) and bits-per-pixel (BPP) averaged on a subset of validation images... Figure 5 presents the rate-distortion and rate-perception curves of PSC compared to several established methods... |
| Researcher Affiliation | Academia | Noam Elata EMAIL Department of Electrical and Computer Engineering Technion Israel Institute of Technology |
| Pseudocode | Yes | Algorithm 1 A single iterative step of row selection Denoted as Select New Rows (H0:k, y0:k, r) Require: Previous sensing rows H0:k, corresponding measurements y0:k, number of new measurements r 1: {xi}s i=1 p(x|y0:k, H0:k) 2: {xi}s i=1 {xi 1 s Ps j=1 xj}s i=1 3: H Append top r singular vectors of x1, . . . , xs 4: return H |
| Open Source Code | Yes | Code implementation for PSC is available at https://github.com/noamelata/PSC. |
| Open Datasets | Yes | We evaluate PSC s effectiveness against established compression methods on the Image Net dataset (Deng et al., 2009)... We use images from the CLIC (Toderici et al., 2020) and DIV2K (Agustsson & Timofte, 2017) to compare Latent-PSC... |
| Dataset Splits | Yes | We compare distortion (PSNR) and bits-per-pixel (BPP) averaged on a subset of validation images, using one image from each of the 1000 classes, following (Pan et al., 2021). |
| Hardware Specification | No | The paper mentions 'high computational complexity, requiring approximately 10,000 NFEs for both compression and decompression (see runtime details in App. A.2)' but does not specify any hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions using 'DDRM (Kawar et al., 2022a)' and 'Range Encoding implemented using (Bamler, 2022)' but does not provide specific version numbers for these or other software libraries/environments. |
| Experiment Setup | Yes | In our implementation we focus on an unsophisticated quantization approach, reducing the precision of y from float32 to float8 (Micikevicius et al., 2022)... Specifically, we find that ΠGDM (Song et al., 2023) produces images with highest photorealism, while DDRM (Kawar et al., 2022a) leads to the lowest distortion. We present both restoration solutions as PSC Perception and PSC Distortion accordingly... Additionally, we condition all posterior sampling steps on a textual description, which must be given along with the original image or inferred using an image captioning module (Vinyals et al., 2016; Li et al., 2022; 2023). |