Research map / Vision-language and multimodal / CLIP and zero-shot
A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks
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.
Proposes SFID, a training-free debiasing method that prunes top RandomForest-important features of frozen VLM representations and imputes them with the mean of low-confidence validation samples, evaluated for gender bias across classification, retrieval, captioning, and generation.
Problem
Gender and racial biases in vision-language models skew outputs across zero-shot classification, text-to-image retrieval, image captioning, and text-to-image generation, while existing debiasing methods are narrow in modality/task or require retraining.
Approach
Use RandomForest feature importance to find attribute-related dimensions of frozen embeddings, and replace them with the average of low-confidence (ambiguous) samples' values, so the embedding stays in-distribution and dimension-preserving for encoders and decoders without retraining.
Key results
- On FACET zero-shot classification with CLIP ResNet-50, SFID reduces demographic parity difference to 9.63 from a baseline of 11.08 (accuracy 50.93 vs 51.87).
- On Flickr30K text-to-image retrieval with CLIP ViT-B/32, SFID lowers Skew@100 to 0.0744 from 0.1721 while keeping R@1 near baseline (58.53 vs 58.91).
- On SDXL text-to-image generation, SFID (HC) reduces the gender-prompt overall mismatch rate to 0.84% from 2.35% baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SFID feature pruning and imputation | CLIP ResNet50 · FACET | Δ DP: 11.08$\pm$0.90 -> 9.63$\pm$0.86 | Helped |
| apply SFID feature pruning and imputation | CLIP ViT-B/32 · FACET | Δ DP: 11.60$\pm$0.93 -> 10.15$\pm$0.85 | Helped |
| apply SFID feature pruning and imputation | XVLM · FACET | Δ DP: 11.72$\pm$0.72 -> 9.91$\pm$0.92 | Helped |
| apply SFID feature pruning and imputation | CLIP ResNet50 · Flickr30K | Skew@100: 0.1883$\pm$0.0939 -> 0.1414$\pm$0.0955 | Helped |
| apply SFID feature pruning and imputation | CLIP ViT-B/32 · Flickr30K | Skew@100: 0.1721$\pm$0.0992 -> 0.0744$\pm$0.0616 | Helped |
| apply SFID feature pruning and imputation | XVLM · Flickr30K | Skew@100: 0.2355$\pm$0.1425 -> 0.2032$\pm$0.1229 | Helped |
| apply SFID feature pruning and imputation | BLIP · MS-COCO | Composite misclassification rate: 2.26$\pm$1.26 -> 1.88$\pm$1.31 | Helped |
| replace zero filling or Gaussian noise with low-confidence imputation (ablation) | SFID on XVLM · Flickr30K | Skew@100: 0.2231 -> 0.0408; 0.4463 -> 0.0408 | Helped |
| replace zero filling or Gaussian noise with low-confidence imputation (ablation) | SFID on XVLM · FACET | Accuracy: 54.66 -> 53.65; 42.85 -> 53.65 | Helped |
| set confidence threshold τ=0.9 rather than τ=0.7 (ablation) | SFID with LCI on XVLM · FACET | Δ DP: 8.11 -> 8.07 | Helped |
| vary confidence threshold from τ=0.7 to τ=0.6, τ=0.8, or τ=0.9 (ablation) | SFID with LCI on XVLM · Flickr30K | R@10: 96.76 -> 96.76; 96.76 -> 96.76; 96.76 -> 96.70 | No clear effect |
Surprising findings
- DeAR with SDXL produces nearly one gender only, raising mismatch rate to 44.64% and Skew on neutral prompts to 99.88.
- Gaussian noise imputation drops XVLM zero-shot accuracy to 42.85 versus 55.70 base, showing out-of-distribution imputation harms utility.
Limitations
- Gains on some fairness metrics are modest and confidence intervals overlap baselines (e.g., CoDi neutral Skew remains high).
- Racial-bias experiments mix FairFace race labels for debiasing with FACET skin-tone labels for evaluation, a mismatch in attributes.
- Hyperparameters k and tau were chosen on XVLM and applied across models; baselines use default repository settings.
Experimental setup
Datasets: FairFace Bias-in-Bios FACET Flickr30K MS-COCO
Benchmarks: FACET Flickr30K MS-COCO FairFace
Metrics: accuracy demographic parity difference Recall@K Skew@100 METEOR SPICE composite misclassification rate mismatch rate Skew
Baselines: DeAR CLIP-clip Prompt-Debias CLIP-CAP BLIP SDXL CoDi
Builds on: CLIP-clip DeAR Prompt-Debias
Closest papers by meaning
- BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS 2024 · similarity 0.90)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.89)
- Interpretable Debiasing of Vision-Language Models for Social Fairness (CVPR 2026 · similarity 0.88)
- Bridge the Modality and Capability Gaps in Vision-Language Model Selection (NeurIPS 2024 · similarity 0.88)
- Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models (ICLR 2025 · similarity 0.88)
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 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.