Causal audit of prompt-conditioned all-in-one image restoration: code and result artifacts
Description
Reproducibility package for the manuscript "Are Degradation Prompts Actually Used? A Cross-Architecture Causal Audit of Prompt-Conditioned All-in-One Image Restoration" (Signal Processing: Image Communication, submitted). The study asks whether the degradation prompts of all-in-one restoration models actually change the output, or whether the models restore about as well without them. Three public checkpoints (PromptIR, InstructIR and CPLIR) were audited on a grid of 68 BSD68 scenes crossed with every task and severity level, and the main interventions were repeated on Rain100L and SOTS. No model was retrained; the only fitted components are linear probes applied to the recorded representations. Contents of SPIC_research_data_v1: - results/: 24 JSON files with the statistics behind every table and figure; per-image intervention samples (raw_samples*.npz, one Delta-PSNR, Delta-SSIM and Delta-LPIPS value per image and condition, with the BSD68 index of each entry, for three degradation seeds); the representations the probes were fitted on (reps.npz: 1292 rows for PromptIR, 816 for InstructIR, 228 for the 12-image CPLIR probe set, covering prompts, pre-prompt features, gate activations and text embeddings); and one UMAP embedding. - experiments/: 20 scripts that regenerate results/, including the figure scripts. Four of them only read results/, so the reported statistics can be checked without a GPU or the model checkpoints. - model_code/: model definitions copied from the official PromptIR and InstructIR releases (InstructIR contains NAFNet blocks), unmodified. - figures/: the five figures as they appear in the manuscript. - docs/: field names and array shapes for the JSON and npz files, and a table mapping each manuscript table to its source file. - MANIFEST.csv: size and SHA-256 of every file. - README.md and requirements.txt: environment, run order and download sources. Statistics are image-clustered paired means (per-image means first, then a t interval over images) with TOST equivalence tests at 0.1, 0.25 and 0.5 dB; probes use image-grouped cross-validation with label-permutation nulls. Recomputing the paired statistics from raw_samples*.npz reproduces paired_tost.json. Numeric arrays in the npz files are stored as decimal strings to keep the archive exact; cast with .astype(float). Benchmark images and model checkpoints are third-party material and are not redistributed here; README.md lists where to download them. Licence: CC BY 4.0 for the authors' scripts, result files and documentation; model_code/ remains under the licences of the original releases.