Research map / Image generation / Text-to-image diffusion
Obliviate: Erasing Concepts from Autoregressive Image Generation Models
Hossein Shakibania, Jonas Henry Grebe, Tobias Braun, Ege Aktemur, Saleh Aslani, Mehmet Görkem Yiğit, Marcus Rohrbach
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 KL distribution matching instead of cross entropy token supervision (ablation) | Obliviate (Liquid) · Explicit content RAB | CDR: 5.26 -> 3.15 | Improved |
| use KL distribution matching instead of cross entropy token supervision (ablation) | Obliviate (Liquid) · MJHQ-30K utility | FID: 23.24 -> 15.41 | Improved |
| use comprehensive concept prompt instead of simple prompt for explicit content (ablation) | Obliviate (Emu3-Gen) · Explicit content | CDR: 27.78 -> 4.76 | Improved |
| use comprehensive concept prompt instead of simple prompt for explicit content (ablation) | Obliviate (Liquid) · Explicit content | CDR: 6.36 -> 3.74 | Improved |
| use comprehensive descriptive prompt instead of simple brand name (ablation) | Obliviate (Emu3-Gen) · Coca-Cola unbranding | CDR: 4.14 -> 91.19 | Hurt |
| use comprehensive descriptive prompt instead of simple brand name (ablation) | Obliviate (Janus-Pro) · Coca-Cola unbranding | CDR: 0.18 -> 19.60 | Hurt |
| increase negative guidance scale eta from 1.0 to 2.0 (ablation) | Obliviate (Emu3-Gen) · Explicit content RAB | CDR: 11.57 -> 5.26 | Improved |
| increase negative guidance scale eta from 1.0 to 2.0 (ablation) | Obliviate (Emu3-Gen) · MJHQ-30K utility | FID: 15.33 -> 25.62 | Hurt |
| increase negative guidance scale eta from 1.0 to 10.0 (ablation) | Obliviate (Janus-Pro) · Explicit content T2I-RiskyPrompt | CDR: 33.48 -> 18.11 | Improved |
| Obliviate versus supervised fine-tuning for brand erasure | SFT (Janus-Pro) · Coca-Cola unbranding | CDR: 32.19 -> 0.18 | Improved |
| Obliviate versus EAR on Janus-Pro explicit content | EAR · MJHQ-30K utility | FID: 31.63 -> 12.31 | Improved |
| Obliviate versus negative prompting for gory content | Original model (Liquid) · Gory content T2I-RiskyPrompt | CDR: 87.65 -> 39.06 | Improved |
Closest papers by meaning
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.87)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.86)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.86)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.86)
- Memories of Forgotten Concepts (CVPR 2025 · similarity 0.86)
- EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.86)
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.
