Visual Question Answering on GQA (Corruption Robustness: Noise, Blur, Weather, Digital)
60.4Accuracy (Noise)Latent Denoising
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Latent DenoisingBackbone=LLaVA+CLIP, Corruption Severity=52026.04 | 60.4 | 59.7 | 62 | 61.6 | |
| BaselineBackbone=LLaVA+CLIP, Corruption Severity=52026.04 | 57.4 | 56.4 | 59 | 58.4 | |
| Latent DenoisingArchitecture=Qwen-2.5-VL, Severity=52026.04 | 53.7 | 57.9 | 62.3 | 61 | |
| BaselineArchitecture=Qwen-2.5-VL, Severity=52026.04 | 47.3 | 51.1 | 54.7 | 52.6 |