# Image Classification on ImageNet Robustness Variants (Adversarial, Rendition, Sketch) V2 (test)

Source: [Image Classification on ImageNet Robustness Variants (Adversarial, Rendition, Sketch) V2 (test)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test>)

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

Data updated: 2026-05-07T08:32:21.891Z

Evaluation of standard Image Classification accuracy and robustness across specific ImageNet variants (V2, Adversarial, Rendition, Sketch) measured on the test split.

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 Robustness Variants (Adversarial, Rendition, Sketch) V2 (test) |
| Metric | Accuracy (ID) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 14 of 14 |
| 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 (ID)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(ID)>): Higher is better; unit: PERCENT.

- [Accuracy (ImageNet V2)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(ImageNet%20V2)>): Higher is better; unit: PERCENT.

- [Accuracy (Adversarial)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(Adversarial)>): Higher is better; unit: PERCENT.

- [Accuracy (Rendition)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(Rendition)>): Higher is better; unit: PERCENT.

- [Accuracy (Sketch)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(Sketch)>): Higher is better; unit: PERCENT.

- [Accuracy (OOD Avg)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(OOD%20Avg)>): Higher is better; unit: PERCENT.

- [Accuracy (Overall Avg)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-robustness-variants-adversarial-rendition-sketch-v2-test.md?metric=Accuracy%20(Overall%20Avg)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Accuracy (ID) | Result date | Report paper |
| --- | --- | --- | --- |
| Adam-SPD | 84.21 | Not recorded | 68fa75841e209131cea6855b |
| FTP | 84.19 | Not recorded | 68fa75841e209131cea6855b |
| SAM | 83.67 | Not recorded | 68fa75841e209131cea6855b |
| Vanilla FT | 83.66 | Not recorded | 68fa75841e209131cea6855b |
| L2-SP | 83.44 | Not recorded | 68fa75841e209131cea6855b |
| LP-FT | 82.99 | Not recorded | 68fa75841e209131cea6855b |
| WiSE-SPD | 81.7 | Not recorded | 68fa75841e209131cea6855b |
| WiSE-FT | 80.94 | Not recorded | 68fa75841e209131cea6855b |
| Linear Prob. | 78.25 | Not recorded | 68fa75841e209131cea6855b |
| SpecPL | 72.12 | Not recorded | 69fc1cac5497b3b670c8861e |
| SpecPL | 72.07 | Not recorded | 69fc1cac5497b3b670c8861e |
| MMRL | 72.03 | Not recorded | 69fc1cac5497b3b670c8861e |
| CoOp | 71.51 | Not recorded | 69fc1cac5497b3b670c8861e |
| Zero-Shot | 67.68 | Not recorded | 68fa75841e209131cea6855b |

### Adam-SPD

Accuracy (ID): 84.21. Result ID: 6993cd143a6c31484918a211.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Implementation](<https://github.com/GT-RIPL/Selective-Projection-Decay.git>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Fine-Tuning","Regularization":"Selective Projection Decay"}. This record does not assert independent reproduction.

### FTP

Accuracy (ID): 84.19. Result ID: 6993cd143a6c31484918a20f.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71051e209131cea4ea0c.md>)

[Implementation](<https://github.com/GT-RIPL/FTP.git>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Fine-Tuning","Regularization":"FTP"}. This record does not assert independent reproduction.

### SAM

Accuracy (ID): 83.67. Result ID: 6993cd143a6c31484918a210.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Fine-Tuning","Optimizer":"SAM"}. This record does not assert independent reproduction.

### Vanilla FT

Accuracy (ID): 83.66. Result ID: 6993cd143a6c31484918a20b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

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

### L2-SP

Accuracy (ID): 83.44. Result ID: 6993cd143a6c31484918a20e.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa64671e209131cea046e0.md>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Fine-Tuning","Regularization":"L2-SP"}. This record does not assert independent reproduction.

### LP-FT

Accuracy (ID): 82.99. Result ID: 6993cd143a6c31484918a20d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6bdf1e209131cea30c50.md>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Linear Probing then Fine-Tuning"}. This record does not assert independent reproduction.

### WiSE-SPD

Accuracy (ID): 81.7. Result ID: 6993cd143a6c31484918a213.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

[Implementation](<https://github.com/GT-RIPL/Selective-Projection-Decay.git>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Weight Interpolation","Regularization":"Selective Projection Decay"}. This record does not assert independent reproduction.

### WiSE-FT

Accuracy (ID): 80.94. Result ID: 6993cd143a6c31484918a212.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

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

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

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

### Linear Prob.

Accuracy (ID): 78.25. Result ID: 6993cd143a6c31484918a20c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

Recorded evaluation configuration: {"Backbone":"CLIP ViT-Base","Protocol":"Linear Probing"}. This record does not assert independent reproduction.

### SpecPL

Accuracy (ID): 72.12. Result ID: 69fc4e0947f34ce924df0389.

[Reporting paper](<https://www.sota2.com/research/paper/69fc1cac5497b3b670c8861e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fc1cac5497b3b670c8861e.md>)

Recorded evaluation configuration: {"baseline":"CoOp"}. This record does not assert independent reproduction.

### SpecPL

Accuracy (ID): 72.07. Result ID: 69fc4e0947f34ce924df038b.

[Reporting paper](<https://www.sota2.com/research/paper/69fc1cac5497b3b670c8861e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fc1cac5497b3b670c8861e.md>)

Recorded evaluation configuration: {"baseline":"MMRL"}. This record does not assert independent reproduction.

### MMRL

Accuracy (ID): 72.03. Result ID: 69fc4e0947f34ce924df038a.

[Reporting paper](<https://www.sota2.com/research/paper/69fc1cac5497b3b670c8861e.md>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### CoOp

Accuracy (ID): 71.51. Result ID: 69fc4e0947f34ce924df0388.

[Reporting paper](<https://www.sota2.com/research/paper/69fc1cac5497b3b670c8861e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6d8b1e209131cea3a814.md>)

[Implementation](<https://github.com/KaiyangZhou/CoOp>)

Recorded evaluation configuration: Not recorded. This record does not assert independent reproduction.

### Zero-Shot

Accuracy (ID): 67.68. Result ID: 6993cd143a6c31484918a20a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa75841e209131cea6855b.md>)

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

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