Researchers fine-tune AI model InVitroVision to read embryo health from time-lapse videos — outperforms ChatGPT on IVF assessment with just 1,000 training images

A research team fine-tuned PaliGemma-2 (a vision-language model that reads images and generates descriptions) on 1,000 embryo time-lapse photos with captions describing cell development, morphology, and stage. The resulting model, InVitroVision, scored higher accuracy than ChatGPT 5.2 and generic base models when asked to describe embryo health and developmental progress.

Unlike existing AI tools for IVF that require large annotated datasets, InVitroVision learns effectively from minimal data—just 1,000 images—because it leverages foundational models trained on billions of images beforehand. This means clinics with smaller embryo image archives can now deploy AI grading without collecting years of labeled examples first.

Fertility clinics currently rely on embryologist judgment calls (which vary between professionals) to select the healthiest embryos for implantation. InVitroVision could standardize these assessments and reduce subjectivity, making embryo selection more consistent across clinics—which matters directly to success rates and patient outcomes in IVF cycles.