Research map / Image generation / Text-to-image diffusion
FailureAtlas: Mapping the Failure Landscape of T2I Models via Active Exploration
Muxi Chen, Zhaohua Zhang, Chenchen Zhao, Mingyang Chen, Wenyu Jiang, Tianwen Jiang, Jianhuan Zhuo, Yu Tang, Qiuyong Xiao, Jihong Zhang, Qiang Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add rule-based pruning that skips extensions of failed nodes (ablation) | FailureAtlas tree search · SD1.5, layer 3 | search space size: reduced to 4.2% at layer 3, 0.4% at extrapolated layer 4 | Improved |
| add prediction-based prioritization of likely-failing nodes (ablation) | FailureAtlas tree search · SD1.5, layer 3 | failure discovery speed: roughly 2x speed-up | Improved |
| pruning applied to a weaker model (SD1.5 vs SDXL Turbo) | rule-based pruning · SD1.5 vs SDXL Turbo | search space reduction: 439,467 vs 1,040,415 explored nodes | Improved |
| use online-trained predictor of node success rate | prioritizer · SD1.5 node success prediction | L1 loss on unseen nodes | Improved |
| increase repeated generations per node beyond 25 (ablation) | node success rate estimate · SD1.5 and SDXL Turbo | average change in success rate: less than 0.6% average change | No clear effect |
| low training-data frequency of an entity or attribute | SD1.5 error slices · LAION-2B-en, layer 1 | error attribution to data scarcity | Mixed |
| add attributes to an entity prompt (layer 1 to layer 2) | SD1.5 and SDXL Turbo · error density by layer | error density: SD1.5 21.3% -> 84.1%; SDXL Turbo 12.0% -> 72.0% | Hurt |
Closest papers by meaning
- Failures Are Fated, But Can Be Faded: Characterizing and Mitigating Unwanted Behaviors in Large-Scale Vision and Language Models (ICML 2024 · similarity 0.87)
- Discovering Failure Modes of Text-guided Diffusion Models via Adversarial Search (ICLR 2024 · similarity 0.86)
- PAGER: Accurate Failure Characterization in Deep Regression Models (ICML 2024 · similarity 0.85)
- AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.85)
- A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective (NeurIPS 2025 · similarity 0.84)
- Adaptive auditing of AI systems with anytime-valid guarantees (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
