# Fair Image Retrieval on CelebA (test)

Source: [Fair Image Retrieval on CelebA (test)](<https://www.sota2.com/research/sota/fair-image-retrieval-on-celeba-test>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-02-26T05:47:44.824Z

Evaluation of fair image retrieval performance on the test split of the CelebA dataset, focusing on fairness metrics like KL Divergence and group skewness.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Fair Image Retrieval |
| Dataset (including any stated split) | CelebA (test) |
| Metric | KL Divergence |
| Unit | SCALAR |
| Direction | Lower is better |
| Results loaded | 9 of 9 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [KL Divergence](<https://www.sota2.com/research/sota/fair-image-retrieval-on-celeba-test.md?metric=KL%20Divergence>): Lower is better; unit: SCALAR.

- [Max Skew](<https://www.sota2.com/research/sota/fair-image-retrieval-on-celeba-test.md?metric=Max%20Skew>): Lower is better; unit: PERCENT.

- [Worst Group AUC](<https://www.sota2.com/research/sota/fair-image-retrieval-on-celeba-test.md?metric=Worst%20Group%20AUC>): Higher is better; unit: PERCENT.

## Recorded results

| Method | KL Divergence | Result date | Report paper |
| --- | --- | --- | --- |
| BEND-VLM | 0.011 | Not recorded | 68fa759e1e209131cea68fe9 |
| BEND-VLM | 0.016 | Not recorded | 68fa759e1e209131cea68fe9 |
| Orth-Cal. | 0.059 | Not recorded | 68fa759e1e209131cea68fe9 |
| DebiasCLIP | 0.066 | Not recorded | 68fa759e1e209131cea68fe9 |
| Orth-Cal. | 0.067 | Not recorded | 68fa759e1e209131cea68fe9 |
| Orth-Proj. | 0.071 | Not recorded | 68fa759e1e209131cea68fe9 |
| Baseline CLIP | 0.118 | Not recorded | 68fa759e1e209131cea68fe9 |
| Baseline CLIP | 0.14 | Not recorded | 68fa759e1e209131cea68fe9 |
| Orth-Proj. | 0.146 | Not recorded | 68fa759e1e209131cea68fe9 |

### BEND-VLM

KL Divergence: 0.011. Result ID: 6993f3423a6c31484919b925.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Implementation](<https://github.com/waltergerych/bend_vlm>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-L-P14"}. This record does not assert independent reproduction.

### BEND-VLM

KL Divergence: 0.016. Result ID: 6993f3423a6c31484919b921.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Implementation](<https://github.com/waltergerych/bend_vlm>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-B-P16"}. This record does not assert independent reproduction.

### Orth-Cal.

KL Divergence: 0.059. Result ID: 6993f3423a6c31484919b91f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f531e209131cea4511d.md>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-B-P16"}. This record does not assert independent reproduction.

### DebiasCLIP

KL Divergence: 0.066. Result ID: 6993f3423a6c31484919b920.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6dbb1e209131cea3badd.md>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-B-P16"}. This record does not assert independent reproduction.

### Orth-Cal.

KL Divergence: 0.067. Result ID: 6993f3423a6c31484919b924.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f531e209131cea4511d.md>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-L-P14"}. This record does not assert independent reproduction.

### Orth-Proj.

KL Divergence: 0.071. Result ID: 6993f3423a6c31484919b91e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f531e209131cea4511d.md>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-B-P16"}. This record does not assert independent reproduction.

### Baseline CLIP

KL Divergence: 0.118. Result ID: 6993f3423a6c31484919b922.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69581e209131cea21ad0.md>)

[Implementation](<https://github.com/OpenAI/CLIP>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-L-P14"}. This record does not assert independent reproduction.

### Baseline CLIP

KL Divergence: 0.14. Result ID: 6993f3423a6c31484919b91d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa69581e209131cea21ad0.md>)

[Implementation](<https://github.com/OpenAI/CLIP>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-B-P16"}. This record does not assert independent reproduction.

### Orth-Proj.

KL Divergence: 0.146. Result ID: 6993f3423a6c31484919b923.

[Reporting paper](<https://www.sota2.com/research/paper/68fa759e1e209131cea68fe9.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f531e209131cea4511d.md>)

Recorded evaluation configuration: {"Model":"CLIP-ViT-L-P14"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.