Research map / Image generation / Text-to-image diffusion
UnHype: CLIP-Guided Hypernetworks for Dynamic LoRA Unlearning
Piotr Wójcik, Maksym Petrenko, Wojciech Gromski, Przemysław Spurek, Maciej Zieba
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLIP-conditioned hypernetwork (UnHype) instead of per-concept ESD-u for object erasure | SD v1.4 object erasure · CIFAR-10 object erasure (avg of 3 classes) | H_o: 86.96 -> 93.85 | Improved |
| Apply UnHype to nudity erasure instead of leaving SD v1.4 unmodified | SD v1.4 · I2P NudeNet total | total detections: 743 -> 8 | Improved |
| Apply UnHype nudity erasure on Flux instead of unmodified model | Flux.1 [dev] · I2P NudeNet total | total detections: 605 -> 32 | Improved |
| Use UnHype instead of EraseAnything for Flux nudity erasure | Flux.1 [dev] · I2P NudeNet total | total detections: 199 -> 32 | Improved |
| Use UnHype for single-model 100-celebrity erasure instead of TRCE | SD v1.4 · 100-celebrity erasure | H_o: 89.85 -> 92.48 | Improved |
| Use UnHype instead of AdvUnlearn against UnlearnDiffAtk | SD v1.4 nudity · UnlearnDiffAtk ASR | attack success rate (%): 3.40 -> 0.00 | Improved |
| Use UnHype instead of AdvUnlearn against Ring-A-Bell | SD v1.4 nudity · Ring-A-Bell ASR | attack success rate (%): 0.00 -> 1.05 | Hurt |
| Use UnHype instead of ESD-u for 10-class Imagenette removal | SD v1.4 · Imagenette | CLIP score: 22.52 -> 30.61 | Improved |
| Use UnHype instead of ScaPre on ImageNet-Confuse5 | SD v1.4 · ImageNet-Confuse5 | Overall Acc: 84.3 -> 87.0 | Improved |
| Use aggressive erasers UCE/RECE instead of UnHype on Confuse5 | SD v1.4 · ImageNet-Confuse5 | Preserve Acc: 83.3 -> 5.6 | Hurt |
| Increase LoRA rank from 1 to 9 (ablation) | UnHype · Imagenette | total ResNet-50 accuracy: 0.06 -> 0.04 | No clear effect |
| Increase LoRA rank from 1 to 12 (ablation) | UnHype · Imagenette | CLIP score: 30.61 -> 30.61 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SAeUron — SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders (ICML 2025) · better on 1 of 1 shared result
- EraseAnything — EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LoRA Unlearns More and Retains More (Student Abstract) (AAAI 2025 · similarity 0.88)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.88)
- Text-to-LoRA: Instant Transformer Adaption (ICML 2025 · similarity 0.88)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.88)
- Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering (ICLR 2025 · similarity 0.88)
- Boosting Alignment for Post-Unlearning Text-to-Image Generative Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
