Context-aware visual redaction with VLMs and inpainting
We built Context-Aware Visual Redaction Pipeline: Leveraging Vision-Language Models for High-Fidelity Content Inpainting, a demo at ICMR 2026.
Training data and model outputs are full of things that should not stay there: PII, logos, regulated or objectionable content. The pipeline detects the target, asks a VLM for a context-aware prompt of the surrounding scene, then inpaints the mask so the result still looks like a real image or video frame. The same stack is useful twice: sanitizing datasets before training, and as a post-processing guardrail on text-to-image and text-to-video generators.