Visual Question Answering on TextVQA (Corruption Robustness Metrics)
36.5Accuracy (Noise Corruption)Latent Denoising
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Latent DenoisingArchitecture=Qwen-2.5-VL, Severity=52026.04 | 36.5 | 35.9 | 73 | 60.5 | |
| Latent DenoisingBackbone=LLaVA+CLIP, Corruption Severity=52026.04 | 36.5 | 26.5 | 40.3 | 38.7 | |
| BaselineBackbone=LLaVA+CLIP, Corruption Severity=52026.04 | 35 | 24.8 | 38.1 | 37.5 | |
| BaselineArchitecture=Qwen-2.5-VL, Severity=52026.04 | 34.1 | 33.9 | 70.9 | 58.9 |