Research map / Image generation / Text-to-image diffusion
Fairness and bias in generative models
Measure and mitigate bias in text-to-image generation.
52 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Text-to-image diffusion, keyword cluster “t2i · preference · alignment”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
15 papers in 2023–24 and 37 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.8% to 4.3%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2023 · Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery (NeurIPS 2023 · 40 citations)
- 2024 · Constructing Dreams Using Generative AI (AAAI 2024 · 61 citations)
- 2025 · T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · 8 citations)
- 2026 · Test-Time Preference Optimization for Image Restoration (AAAI 2026 · 1 citation)
Earliest papers here
- Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery (NeurIPS 2023 · 40 citations)
- Stable Bias: Evaluating Societal Representations in Diffusion Models (NeurIPS 2023 · 18 citations)
- On Measuring Fairness in Generative Models (NeurIPS 2023 · 2 citations)
What has been tried
29 papers here report 265 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 9%, hurt 23%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Do not train the text encoder during LoRA fine-tuning | CelebA-HQ · FID | 119.8 -> 157.2 | Mixed | Can Protective Perturbation Safeguard Personal Data from Being Exploited by Stable Diffusion? (CVPR 2024 · 15 citations) |
| Increase retrieval count from 1 to 3 | TIFA (Avg) · TIFA score | 83.17 -> 82.57 | Hurt | CPR: Retrieval Augmented Generation for Copyright Protection (CVPR 2024 · 14 citations) |
| Fine-tune on T2ISafety data | ImageGuard test set · Overall F1 | 0.551 -> 0.840 | Helped | T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · 8 citations) |
| Remove debiased query from FairRAG | Human generation prompts · Intersectional diversity | 0.438 -> 0.414 | Hurt | FairRAG: Fair Human Generation via Fair Retrieval Augmentation (CVPR 2024 · 7 citations) |
| Flip 5% of feedback labels | Robotic arm, all solutions · QD score | 72.5 ± 0.9 -> 72.4 ± 0.8 | No clear effect | Quality Diversity through Human Feedback: Towards Open-Ended Diversity-Driven Optimization (ICML 2024 · 4 citations) |
| Replace full fine-tuning with selective fine-tuning | UTKFace (Female; protected White) · WST | 83.00 -> 84.26 | Helped | AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · 2 citations) |
| Set the similarity-loss threshold to ε = 0.5 | Combinatorial object generation · concept integration | Helped | Redefining <Creative> in Dictionary: Towards an Enhanced Semantic Understanding of Creative Generation (CVPR 2025 · 2 citations) | |
| Add sparse regularization to LinEAS, reducing intervention support to about 1% | RTP and TET toxicity mitigation · toxicity | No clear effect | LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss (NeurIPS 2025 · 2 citations) | |
| Add adaptive feature selection to the average bias direction | COCO human captions, negative bias · More negative rate | 82.2% -> 80.2% | Hurt | Implicit Bias Injection Attacks against Text-to-Image Diffusion Models (CVPR 2025 · 1 citation) |
| Use probabilistic modeling instead of numerical modeling | Simulated ranking from scratch, no preference noise · MSE | about 1500 comparisons | Helped | K-Sort Arena: Efficient and Reliable Benchmarking for Generative Models via K-wise Human Preferences (CVPR 2025 · 1 citation) |
| Replace ResNet18 shadow model with ConvNeXt-B | CIFAR10 · accuracy | 0.997 -> 0.995 | No clear effect | Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training? (CVPR 2025 · 1 citation) |
| Apply amplification to Anti-NAF prompts | POKEMON · FAR@5%AR | 8.52% -> 77.36% | Helped | VA3: Virtually Assured Amplification Attack on Probabilistic Copyright Protection for Text-to-Image Generative Models (CVPR 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 17 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CelebA · 3 of 17 papers
- COCO · 2 of 17 papers
- FairFace · 2 of 17 papers
Baselines
- FairDiffusion · 2 of 17 papers
- UCE · 2 of 17 papers
- DALL-E 3 · 2 of 17 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 52 papers; from OpenAlex and the papers' affiliations).
- The Chinese University of Hong Kong · 3 papers
- Massachusetts Institute of Technology · 2 papers
- Amazon · 2 papers
- Chinese Academy of Sciences · 2 papers
- Harvard University · 2 papers
- Fudan University · 2 papers
- The University of Hong Kong · 2 papers
- Huawei · 2 papers
- Shanghai Jiao Tong University · 2 papers
Use it
Download BibTeX of 23 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Inference-time T2I improvement (117 papers)
- T2I evaluation and controllable adapters (149 papers)
- Preference optimization for diffusion (82 papers)
Elsewhere in the topic
- Compositional text-to-image guidance (194 papers)
- Layout-to-image control (164 papers)
- Text-driven synthesis and semantic guidance (137 papers)
- Personalized and region-controlled generation (136 papers)
- Multi-subject personalization (82 papers)
- Concept erasure in diffusion models (76 papers)
- Text generation and structured prediction (65 papers)
- Multi-concept composition (56 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).