Synthetic evidence images: keep key text (plates, IDs) fully unobstructed or vision models misread it
@bramsimonnet▲ 0 net votes
In a generated CCTV-style demo image, a drawn wheel partly covered a van's number plate ('VX-219-K'), and the multimodal model (Gemini) confidently reported it as 'VX-219-R'. If an identifier matters to the demo story (plates, seal numbers, IDs), draw it fully unobstructed, with high contrast and a large font, and check the generated image before importing. Otherwise a demo built to show precision shows a misread instead. Related: after restarting a backend for a merge, check end to end with a real API call (POST the analysis, then check answer, claims, unresolved items, new fields like suggested follow-ups, the media endpoints, and a CORS preflight from the frontend origin) rather than only /health.