Example-based spatial guidance: training-free concept erasure in diffusion models
Example-Based Spatial Guidance for Training-Free Concept Erasure in Diffusion Models
Translation in progress — untranslated text is omitted.
The training-free concept erasure method EBSG provides fine-grained spatial guidance for text-to-image diffusion models using user-editable example packs, decomposing broad unsafe concepts into subconcepts and applying localized steering signals. It also introduces a VLM fine-grained evaluation protocol to prevent model collapse from bypassing safety checks via binary evaluators. EBSG achieves state-of-the-art erasure performance on 4 nudity benchmarks, 7 I2P harmful concept slices, and 4 MJA metaphor categories, supports multi-concept removal, and can transfer to SD3.
- T2I diffusion models generate violating content, while existing erasure methods compress multi-faceted safety standards into a single text embedding for global suppression, and binary detectors (NudeNet, Q16) may also misjudge semantically collapsed noise images as safe.
- EBSG splits broad concepts into sub-concept families with target words, example images, and safety anchors; at inference time it first uses CAS gating to decide when to intervene, then uses text/image dual probes to locate regions, and finally, according to the mask, guides each region from the target direction toward the safety anchor; it also comes with a Qwen3-VL-8B four-category review protocol.
- The authors report that EBSG achieves SR 84.2–97.4% on four nudity benchmarks with the lowest NR, and single-/multi-concept average SR of 73.6%/73.1% on the seven I2P concepts, both higher than training-free baselines; after transfer to SD3, MJA average SR is 51.0% (SAFREE is 32.8%).
- The authors note that per-image time is 5.3× the baseline, anchor drift exists at high guidance scales, concept pack coverage and calibration still require manual work, spatially overlapping concepts have not been fully tested, and repeated experiments are reported only on three nudity benchmarks.
Related papers
Both use: UnlearnDiff
Both use: MMA-Diffusion
Show 3 more papers
Both use: MMA-Diffusion