Erase but Preserve: Controllable Removal of Copyrighted Animation Characters via Optimized Semantic Anchors
| Source: arXiv AI
Tags: diffusion models, copyright, concept erasure, text-to-image, ACM MM, AI safety
Accepted at ACM MM 2026, this method erases specific copyrighted animation characters from diffusion models by optimizing a surrogate anchor embedding — achieving state-of-the-art erasure with tunable strength, multi-target support, and compatibility with existing model-modification baselines.
Details
Diffusion models can reproduce distinctive animation characters well enough to raise copyright liability, but existing concept erasure methods hit two walls: model modification methods struggle to find anchor representations for the highly distinctive characters, while prompt-based steering lacks the precision needed for reliable erasure without damaging surrounding content.\n\nThis paper works in the continuous textual representation space rather than at model weights or prompts. The method optimizes an anchor embedding that acts as a character surrogate using two constraints: structural (preserving spatial composition) and detailed (preserving fine-grained visual features of nearby concepts). At generation time, a structure-aware adaptive strategy replaces target-related embeddings with the anchor, with controllable erasure degree.\n\nThe approach achieves state-of-the-art erasure effectiveness and image fidelity preservation on standard benchmarks, supports multi-target removal in a single pass, and transfers across base models without re-optimization. The optimized anchors are plug-and-play with existing model-modification baselines, boosting their performance. Accepted at ACM MM 2026, the top multimedia research venue. For companies deploying text-to-image generation commercially, this provides a practical tool for copyright compliance without full model retraining.