# Image Classification on ImageNet OOD A/R/S variants (test)

Source: [Image Classification on ImageNet OOD A/R/S variants (test)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-ood-a-r-s-variants-test>)

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

Data updated: 2026-02-26T05:34:17.968Z

Evaluation of Image Classification performance across the specific test splits of the ImageNet Out-of-Distribution variants: ImageNet-A, ImageNet-R, and ImageNet-S, reporting individual accuracies and the overall average.

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 | Image Classification |
| Dataset (including any stated split) | ImageNet OOD A/R/S variants (test) |
| Metric | Accuracy (ImageNet-A) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 5 of 5 |
| 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

- [Accuracy (ImageNet-A)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-ood-a-r-s-variants-test.md?metric=Accuracy%20(ImageNet-A)>): Higher is better; unit: PERCENT.

- [Accuracy (ImageNet-R)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-ood-a-r-s-variants-test.md?metric=Accuracy%20(ImageNet-R)>): Higher is better; unit: PERCENT.

- [Accuracy (ImageNet-S)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-ood-a-r-s-variants-test.md?metric=Accuracy%20(ImageNet-S)>): Higher is better; unit: PERCENT.

- [Average Accuracy (OOD)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-ood-a-r-s-variants-test.md?metric=Average%20Accuracy%20(OOD)>): Higher is better; unit: PERCENT.

- [OOD Delta (%)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-ood-a-r-s-variants-test.md?metric=OOD%20Delta%20(%25)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Accuracy (ImageNet-A) | Result date | Report paper |
| --- | --- | --- | --- |
| TPGM-C | 59.29 | Not recorded | 68fa6edf1e209131cea428cb |
| TPGM | 58.31 | Not recorded | 68fa6edf1e209131cea428cb |
| WISE | 58.26 | Not recorded | 68fa6edf1e209131cea428cb |
| Zero-Shot | 52.21 | Not recorded | 68fa6edf1e209131cea428cb |
| Vanilla FT | 49.67 | Not recorded | 68fa6edf1e209131cea428cb |

### TPGM-C

Accuracy (ImageNet-A): 59.29. Result ID: 699065053a6c314849003d83.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

[Implementation](<https://github.com/PotatoTian/TPGM>)

Recorded evaluation configuration: {"Pre-trained":"CLIP","Backbone":"ViT-L","Protocol":"Proposed (Controlled)"}. This record does not assert independent reproduction.

### TPGM

Accuracy (ImageNet-A): 58.31. Result ID: 699065053a6c314849003d84.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

[Implementation](<https://github.com/PotatoTian/TPGM>)

Recorded evaluation configuration: {"Pre-trained":"CLIP","Backbone":"ViT-L","Protocol":"Proposed"}. This record does not assert independent reproduction.

### WISE

Accuracy (ImageNet-A): 58.26. Result ID: 699065053a6c314849003d82.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6cd31e209131cea3601a.md>)

[Implementation](<https://github.com/mlfoundations/wise-ft>)

Recorded evaluation configuration: {"Pre-trained":"CLIP","Backbone":"ViT-L","Protocol":"Weight Interpolation"}. This record does not assert independent reproduction.

### Zero-Shot

Accuracy (ImageNet-A): 52.21. Result ID: 699065053a6c314849003d81.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

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

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

Recorded evaluation configuration: {"Pre-trained":"CLIP","Backbone":"ViT-L","Protocol":"Zero-Shot"}. This record does not assert independent reproduction.

### Vanilla FT

Accuracy (ImageNet-A): 49.67. Result ID: 699065053a6c314849003d80.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6edf1e209131cea428cb.md>)

Recorded evaluation configuration: {"Pre-trained":"CLIP","Backbone":"ViT-L","Protocol":"Fine-Tuning"}. This record does not assert independent reproduction.

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