TIQA: Human-Aligned Perceptual Text Quality Assessment in Generated Images
This work introduces Text-in-Image Quality Assessment (TIQA), a new no-reference task to measure how well rendered text looks to humans, separate from whether the text is linguistically correct. The model ANTIQA was developed to predict this perceptual quality score using image features and edge maps. Results show ANTIQA accurately predicts human scores, proving that visual text fidelity is a crucial factor in overall image preference.