SketchVLM: Letting VLMs draw on images while explaining their reasoning
github.comI really like the non-destructive approach to indicating things in an image via SVG. I’ve played around with this using pure generative models (such as Nano Banana), and they're impressively proficient at highlighting or marking specific areas.
Attempts to trip up the training data which is obviously fitted on four-legged animals and it was still highly successful.