Research map / Image generation / Text-to-image diffusion

Multi-Group Proportional Representations for Text-to-Image Models

CVPR 2025 · Image generation › Text-to-image diffusion · approach: Fairness and bias in generative models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Applies the Multi-Group Proportional Representation (MPR) metric to measure intersectional demographic representation in text-to-image models and fine-tunes Stable Diffusion with an MPR-based LoRA objective to reduce representational harm.

Problem

Text-to-image models can misrepresent demographic groups, propagate stereotypes, and under-represent minorities, and existing metrics lack a systematic, theoretically grounded way to measure and control intersectional representational harm.

Approach

Adapts the MPR metric (originally for retrieval) as a supremum over a function class of differences between generated-image and reference-dataset statistics, with flexible reference distributions and function classes (linear or decision trees), and uses it as a fine-tuning objective with buffered c* functions and CLIP/DINO regularization.

Key results

• Fine-tuned model reaches MPR (GAR) 0.25 versus 1.89 for vanilla SD v1.4 and 0.44 for ITI-GEN, averaged over seven professions, with CLIP score 0.30.
• MPR with gender, age, and race decision trees shows baselines such as FairDiffusion and UCE degrade on intersectional groups (FairDiffusion MPR (GAR) 1.32, UCE 1.56).
• Across traits, MPR values of 0.39 to 0.84 are measured for SD v1.4, SD v2.1, and SDXL, with depth-3 trees showing larger values than depth-1.

Measured effects reported in the paper

Surprising findings

• UCE's performance when race is included falls below the vanilla model's MPR, indicating gender-only debiasing can worsen intersectional disparities.
• The historically accurate ENIAC-programmer reference showed FairDiffusion achieves gender parity but still misses the women-only historical distribution.

Limitations

• Relies on attribute classifiers (linear probes on CLIP embeddings trained with FairFace, BLIP2 VQA pseudo-labels) whose accuracy is not fully validated.
• Choice of reference distribution is context-dependent and remains challenging; the paper notes needs for context-appropriate references.
• Some embedding-based identity models from prior work were proprietary and not validated experimentally.

Experimental setup

Datasets: FairFace CDC disability statistics Google Images search results

Benchmarks: FairFace test set (reference) CLIP score on occupation prompts

Metrics: MPR CLIP score total variation via decision trees

Baselines: Vanilla SD v1.4 Entigen FairDiffusion ITI-GEN UCE

Builds on: multicalibration multiaccuracy MPR for retrieval FairDiffusion ITI-GEN

Closest papers by meaning

• Multi-Group Proportional Representation in Retrieval (NeurIPS 2024 · similarity 0.89)
• FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.86)
• OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes (ICLR 2025 · similarity 0.86)
• Association of Objects May Engender Stereotypes: Mitigating Association-Engendered Stereotypes in Text-to-Image Generation (NeurIPS 2024 · similarity 0.86)
• OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.85)
• SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.85)