The authors apply 13 image transformations (vertical flip, rotation, blur, noise, fog, etc.) to ImageNet test images and compare four prompting strategies: no attribute, random attribute, wrong attribute, and correct attribute. Conditioning on the ground-truth attribute consistently improves accuracy over the no-attribute baseline, with an average gain of +1.16% across all transformations. Self-inferred attributes (CLIP's own estimate) yield a similar +1.17% gain. On the full ImageNet test set with ViT-B/16, using the top-4 attribute composition achieves 68.80% versus 66.72% for the single template and 68.32% for 80-template prompt ensembling.
Evidence
correlational
Key metric
average accuracy gain +1.16% (correct z) and +1.17% (self-inferred z) across 13 transformations on ImageNet; ViT-B/16 ImageNet: 68.80% (top-4 attributes) vs 66.72% (single template) vs 68.32% (80 templates)
Caveat
The paper notes sensitivity to text description wording: using different synonyms for the same attribute can have non-trivial effects. Ablation with random strings of equal length shows a smaller but non-zero gain, suggesting some benefit from added tokens.