Research map / Image generation / Text-to-image diffusion
FairSteer: Cross-Attention Steering Towards a Fairer Text-Guided Image Generation
Tatiana Gaintseva, Akshit Achara, Greg Slabaugh, Jiankang Deng, Ismail Elezi
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 |
|---|---|---|---|
| apply EquiSteer cross-attention steering for gender debiasing | SD-1.5 · Gender, neutral prompts (8 professions) | avg parity gap Delta: 0.403 -> 0.051 | Improved |
| EquiSteer vs training-free TEI baseline | SD-1.5 · Gender, neutral prompts | avg parity gap Delta: 0.167 -> 0.051 | Improved |
| add orthogonalisation against attribute directions to adaptive steering (ablation) | EquiSteer · SD-1.5 gender, neutral prompts | neutral Delta: 0.165 -> 0.095 | Improved |
| add orthogonalisation without gate (ablation) | EquiSteer · SD-1.5 gender, male-specific prompts | male-specific ratio: 0.379 -> 0.230 | Hurt |
| add prompt-aware gate (ablation) | EquiSteer (add & erase) · SD-1.5 gender | neutral Delta / male-specific ratio: neutral 0.095 -> 0.051; male-specific 0.230 -> 0.495 | Improved |
| apply EquiSteer to race debiasing | SDXL and SANA · Race, neutral prompts | parity gap Delta: 0.172 -> 0.042 (SDXL), 0.178 -> 0.039 (SANA) | Improved |
| apply EquiSteer to eyeglasses | SANA · Eyeglasses, neutral prompts (BLIP-VQA) | parity gap Delta: 0.234 -> 0.175 | Improved |
| debias four concepts jointly in one pass | EquiSteer · Joint-4 gender+race+age+body type | parity gap Delta: gender SDXL 0.381 -> 0.012; body type SANA 0.391 -> 0.178 | Improved |
| apply EquiSteer to compositional/multi-subject prompts | SANA · Gender, compositional prompts | parity gap Delta: 0.295 -> 0.155 | Improved |
| EquiSteer vs FairImagen on FairImagen setup | SDXL · Gender, pooled Delta (10 steps) | pooled Delta: 0.098 -> 0.008 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FairImagen — FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.89)
- OneActor: Consistent Subject Generation via Cluster-Conditioned Guidance (NeurIPS 2024 · similarity 0.88)
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.88)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.88)
- Fair Text-to-Image Diffusion via Fair Mapping (AAAI 2025 · similarity 0.88)
- FairQueue: Rethinking Prompt Learning for Fair Text-to-Image Generation (NeurIPS 2024 · similarity 0.87)
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.
