Abstract
A pre-generation pipeline intercepts user text prompts for a generative image model and performs intellectual-property-aware prompt rewriting. Protected-entity spans are detected using an entity recognizer covering trademarks, copyrighted characters, trade dress, and celebrity likenesses. Detected spans are embedded and compared against a multi-modal IP vector registry to confirm matches and retrieve a visual attribute vector that encodes visual characteristics associated with the protected entity. The prompt is decomposed into a protected identity layer and a descriptive intent layer. Non-infringing descriptors derived from the retrieved visual attributes are injected into the descriptive intent layer, and a language model generates a rewritten prompt that omits protected identity references. A fidelity validator computes semantic similarity between the original descriptive intent and the rewritten prompt and controls whether the rewritten prompt proceeds to generation or triggers further rewriting or user clarification.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Semantic Prompt Analysis and Intent-Preserving Reconstruction for Generative AI Content Systems", Technical Disclosure Commons, (June 30, 2026)
https://www.tdcommons.org/dpubs_series/10651