Sensitivity Ridges in Conditioned Generative Models - Code and Data Archive (EPIA 2026)

Sensitivity Ridges in Conditioned Generative Models — Data Archive ==================================================================== Companion archive for: Schwind, M., Nuhic, J., Martins, P.: Sensitivity Ridges in Conditioned Generative Models. In: Progress in Artificial Intelligence (EPIA 2026), Springer LNAI (2026). Living repository (code + tool): https://github.com/th-nuernberg/sensitivity-ridges Contents -------- sensitivity_maps_and_eval_data.tar.gz (~10 MB) outputs/ ......... all pre-computed sensitivity maps from the paper: 50x50 DINOv2/CLIP/pixel-MSE grids for Flux.1-schnell (50 triplets), cross-architecture runs (SD 3.5, DreamShaper 8, PixArt-alpha, DeepFloyd IF, Kandinsky 3, SDXL, Hunyuan-DiT), FLUX.2 Klein 9B and FLUX.2-dev (30x30), five SD 1.5 variants, GPT-2 and Qwen2.5 language-model grids, and the anonymized human-evaluation data (2,129 clean judgments, 45 annotators) with the granular re-analysis. data/prompts/ .... the 160-prompt bank and the triplet definitions as JSON: 50 hand-designed triplets (the paper's 25 hand-picked set is triplets 00-24) plus the 25 bank-sampled random triplets (seed 42) that complete the paper's 50-triplet dataset. human_eval_stimulus_images.tar.gz (~179 MB) The 3,500 JPEG stimulus images shown to annotators in the human study (pair definitions in outputs/26_mixed_eval/pairs.json above). code_snapshot.tar.gz (~320 KB) Snapshot of the repository code at the time of archiving (experiment scripts, figure scripts, Ridge Explorer web tool). For the maintained version, use the GitHub repository. Format notes ------------ Sensitivity maps are numpy .npy files: 50x50 (or 30x30) float grids on the barycentric simplex; entries outside the simplex are NaN. Grid point (i,j) corresponds to weights (1-a-b, a, b) with a = i/(G-1), b = j/(G-1). Human-eval responses are one JSONL file per anonymized annotator hash. License: CC BY 4.0 (see DATA_LICENSE.txt). No personal data included.