{"name":"Concept-Brittleness","repo":"MUSE-AI-Lab/Concept-Brittleness","url":"https://midorreal.com/project/muse-ai-lab-concept-brittleness","repository":"https://github.com/MUSE-AI-Lab/Concept-Brittleness","description":"Concept Brittleness","facts":{"language":{"value":"Python","source":"https://github.com/MUSE-AI-Lab/Concept-Brittleness","checked":"21 Nov 2026"},"releases":{"value":"0","source":"https://github.com/MUSE-AI-Lab/Concept-Brittleness/releases","checked":"21 Nov 2026"},"contributors":{"value":"2","source":"https://github.com/MUSE-AI-Lab/Concept-Brittleness/graphs/contributors","checked":"21 Nov 2026"},"archived":{"value":"no","source":"https://github.com/MUSE-AI-Lab/Concept-Brittleness","checked":"21 Nov 2026"}},"stars":243.0,"writeup":{"what_it_is":"This repository contains the code for a method that interprets object-dependent concept brittleness in text-to-image diffusion models. It trains a sparse autoencoder at each denoising step, builds class-level concept prototypes, and applies a prototype-guided correction at inference time to repair failed generations without retraining.","audience":"Researchers studying text-to-image diffusion models and their failure modes","claims":[{"kind":"specific","claim":"Makes failed generations repairable at inference time without retraining","excerpt":"the failures **repaired at inference time** by nudging features toward a class-level concept prototype — no retraining","status":"not_checked"},{"kind":"specific","claim":"Reports improved CLIP-I scores across the five listed backbones","excerpt":"| **SD 3.5 + Ours** | **0.797** | **0.297**","status":"not_checked"},{"kind":"specific","claim":"Correction adds almost no inference time","excerpt":"the correction adds almost no inference time","status":"not_checked"},{"kind":"specific","claim":"Supports five backbones: FLUX.1-dev, SD 3.5, SDXL, PixArt-α and SD 1.5","excerpt":"| `sd1.5` | Stable Diffusion 1.5 |","status":"not_checked"},{"kind":"specific","claim":"Results are reported for a single fixed seed on held-out prompts","excerpt":"Held-out prompts, 256×256, official sampler settings of each backbone, seed 2026","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-16"},"why_now":null,"owner_supplied":null,"score":{"verdict":"real","hype":42,"reality":60,"gap":-18,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-10-28T15:08:18+00:00","text":"Repository created","source":"https://github.com/MUSE-AI-Lab/Concept-Brittleness"}],"cite":"Concept-Brittleness is a GitHub project at https://github.com/MUSE-AI-Lab/Concept-Brittleness written mainly in Python, described there as \"Concept Brittleness\".","attribution":"Data from Mid or Real, https://midorreal.com/project/muse-ai-lab-concept-brittleness, checked 22 November 2026. Attribution with a link is required.","still_real":null}