An open model from the WeChat vision team handles detection, counting, OCR and pixel masking without bolting on…
A five-lab study finds text color and contrast alone can shift vision-language model verdicts, with Qwen2-VL most affected.