Research map / Vision-language and multimodal / Multimodal LLMs
On Fairness of Unified Multimodal Large Language Model for Image Generation
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.
Benchmarks gender and race bias in unified multimodal LLMs for text-to-image generation, localizes it to the language model, and reduces it with synthetic balanced data plus a balanced preference (odds-ratio) loss on VILA-U.
Problem
Unified multimodal LLMs that both understand and generate images may propagate demographic stereotypes in generated images, and it is unclear where this bias originates and how to mitigate it in token-based autoregressive generation.
Approach
A balanced preference loss built on reference-free odds-ratio preference optimization that penalizes differences in preference between demographic groups' generated images, applied after SFT on diffusion-synthesized demographically balanced data; bias is localized via linear probing of the encoder and JSD over image-token distributions of the LM.
Key results
- VILA-U gender bias drops from 0.89 to 0.25 with BPO, with Inception score reported at 2.10 versus 1.87 for the base model (paper cites 12.2% IS increase and 71.91% gender-bias reduction).
- Janus-Pro shows the worst gender bias (0.90) among U-MLLMs, versus 0.67 for Stable Diffusion.
- On race, Finetune (T→I) reaches 0.23 and BPO reaches 0.26 vs 0.48 for base VILA-U; BPO achieves the best G×R bias of 0.15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply balanced preference optimization | VILA-U · Gender debiasing occupation prompts | Gender bias: 0.89 -> 0.25 | Helped |
| Finetune on balanced data in text-to-image fashion | VILA-U · Gender debiasing occupation prompts | Gender bias: 0.89 -> 0.27 | Helped |
| Finetune on balanced data in image-to-text fashion | VILA-U · Gender debiasing occupation prompts | Gender bias: 0.89 -> 0.83 | Helped |
| Apply balanced preference optimization | VILA-U · Race debiasing occupation prompts | Race bias: 0.48 -> 0.26 | Helped |
| Finetune on balanced data in text-to-image fashion | VILA-U · Race debiasing occupation prompts | Race bias: 0.48 -> 0.23 | Helped |
| Apply balanced preference optimization | VILA-U · G×R debiasing occupation prompts | G×R bias: 0.24 -> 0.15 | Helped |
| Finetune on balanced data in text-to-image fashion | VILA-U · G×R debiasing occupation prompts | G×R bias: 0.24 -> 0.17 | Helped |
| Ask for diversity in generated images | VILA-U · Gender debiasing occupation prompts | Gender bias: 0.89 -> 0.56 | Helped |
| Train on demographic images in a discriminative image-to-text fashion | VILA-U · Image generation | Demographic bias | No clear effect |
| Prompt for diversity | TokenFlow · Occupation prompts | Gender bias: 0.84 -> 0.69 | Helped |
| Finetune TokenFlow | TokenFlow · Occupation prompts | Gender bias: 0.84 -> 0.55 | Helped |
| Finetune on balanced data in text-to-image fashion | VILA-U · Gender debiasing occupation prompts | CLIP-IQA: 0.84 -> 0.77 | Hurt |
Surprising findings
- Partial alignment: U-MLLMs refuse explicit demographic VQA answers yet still generate strongly biased images.
- Discriminative finetuning on demographic labels did not reduce generation bias.
Limitations
- Only overt demographic categories (gender, race) studied; age, culture, religion not covered.
- Only two publicly available models with fine-tuning code were evaluated; most U-MLLMs not fully open-source.
- Did not explore alternative preference optimization objectives.
Experimental setup
Datasets: FairFace synthetic demographic-balanced images from FLUX.1-dev occupation prompt dataset
Benchmarks: occupation-based generation prompts (50 prompts, 160 images each)
Metrics: Representation Disparity (gender, race, G×R) CLIP-S CLIP-IQA Inception Score
Baselines: Stable Diffusion Janus Janus-Pro Show-o TokenFlow Emu3 Prompt Engineering Finetune (I→T) Finetune (T→I)
Builds on: DPO ORPO VILA-U FLUX.1-dev LoRA
Closest papers by meaning
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.90)
- HermesFlow: Seamlessly Closing the Gap in Multimodal Understanding and Generation (NeurIPS 2025 · similarity 0.89)
- UniGen: Enhanced Training & Test-Time Strategies for Unified Multimodal Understanding and Generation (NeurIPS 2025 · similarity 0.89)
- SEED-Bench: Benchmarking Multimodal Large Language Models (CVPR 2024 · similarity 0.89)
- OpenING: A Comprehensive Benchmark for Judging Open-ended Interleaved Image-Text Generation (CVPR 2025 · similarity 0.89)
- Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models (AAAI 2025 · similarity 0.89)
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.