Rethinking Visual Provenance: Detection and Watermarking Across Direct Visual Generation and LLM-Driven Code Rendering
A conceptual agenda compares detection and watermarking for generated and code-rendered media.
The paper outlines a production-centered framework comparing provenance detection and watermarking when media comes from generative models or from LLM-written code that is later rendered. It distinguishes passive inference, message recovery, and authenticated provenance, and organizes watermarks by production stage across images, video, plots, SVG, and agent workflows. Interfaces from Claude, OpenAI, and rendering tools ground the discussion. The work poses ten research questions, reports no experiments, and claims no new theorems.
- Compares detection and watermarking for generation versus code rendering.
- Splits verification into inference, recovery, and authenticated provenance.
- Grounded in Claude, OpenAI, and rendering-tool interfaces.
- Poses ten research questions with no experiments or new theorems.
From the original article
1 min read at arxiv.orgAI systems create images and videos with image/video generation models or by writing code and graphics descriptions that are then rendered. These routes can produce similar visible artifacts but expose different representations, intervention points, and provenance evidence. We develop a production-centered framework that compares detection and watermarking across both routes.
By Zheng Gao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.