Research map / Image generation / Text-to-image diffusion
One Image is Worth a Thousand Words: A Usability Preservable Text-Image Collaborative Erasing Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Co-Erasing, a text-image collaborative concept erasing method for text-to-image diffusion that adds self-generated images of the target concept as visual prompts (with text-guided refinement) to ESD-style fine-tuning, improving erasure efficacy while preserving usability.
Problem
Concept erasing in text-to-image diffusion models relies only on manually crafted text prompts, which yields incomplete erasure of entangled concepts and degrades benign generations; the text-image modality gap limits efficacy.
Approach
Describe the target concept jointly by text and by self-generated images of that concept (fed through a decoupled cross-attention branch used only in training), and refine image features with the target word so that only concept-relevant visual features guide negative guidance erasure.
Key results
- On nudity erasure, Co-Erasing reaches pre-ASR 0.85 and ASR 16.96 versus ESD's 20.42 and 76.05, with FID 18.77 and CLIP 0.302 (near ESD's 18.18 and 0.302).
- Ring-A-Bell RAB_K16 drops from 0.47 (ESD) to 0.18 with ESD+Co-Erasing; MACE+Co-Erasing reaches 0.00 on all RAB settings.
- Ablation on nudity: text-only gives ASR 76.05 and FID 17.29; text+image+refinement gives ASR 16.96 and FID 18.77, showing the image branch lowers ASR at some FID cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add image prompts to text-only erasing (ablation) | Co-Erasing on nudity · Nudity | ASR: 76.05 -> 32.60 | Helped |
| Add image prompts to text-only erasing (ablation) | Co-Erasing on nudity · Nudity | FID: 17.29 -> 22.98 | Hurt |
| Add text-guided refinement to image-only erasing (ablation) | Co-Erasing on nudity · Nudity | ASR: 41.52 -> 27.12 | Helped |
| Add text-guided refinement to image-only erasing (ablation) | Co-Erasing on nudity · Nudity | FID: 25.82 -> 24.56 | Helped |
| Combine text, image prompts, and text-guided refinement (ablation) | Co-Erasing on nudity · Nudity | ASR: 32.60 -> 16.96 | Helped |
| Combine text, image prompts, and text-guided refinement (ablation) | Co-Erasing on nudity · Nudity | FID: 22.98 -> 18.77 | Helped |
| Use generated rather than real images as erasing prompts (ablation) | Co-Erasing · Nudity | ASR: 33.90 -> 16.96 | Helped |
| Use generated rather than real images as erasing prompts (ablation) | Co-Erasing · Nudity | FID: 23.89 -> 18.77 | Helped |
| Increase text descriptor count from 1 to 10 words (ablation) | Text-based erasing · Nudity | ASR: 76.05 -> 29.66 | Helped |
| Increase text descriptor count from 1 to 10 words (ablation) | Text-based erasing · Nudity | FID: 17.29 -> 17.83 | Hurt |
| Use Co-Erasing rather than ESD | Stable Diffusion · Nudity | ASR: 76.05 -> 16.96 | Helped |
| Use Co-Erasing rather than ESD | Stable Diffusion · Nudity | FID: 18.18 -> 18.77 | Hurt |
Surprising findings
- Using only image prompts without text-guided refinement raises FID above 24, showing unrefined visual features harm benign generation.
- Self-generated images from the clean model outperform real NSFW dataset images on both efficacy and usability.
Limitations
- Image generation step and reliance on classifier filtering for attack methods add setup cost.
- Some concepts show higher FID than competitors (e.g., garbage truck, church), indicating a usability trade-off remains.
- Reported failure cases and limitations are only briefly described in the appendix.
Experimental setup
Datasets: self-generated images from Stable Diffusion NSFW dataset (nsfw_data_scraper) I2P COCO-10k
Benchmarks: ASR (pre-ASR and adversarial) P4D CCE Ring-A-Bell (RAB) FID on COCO-10k CLIP score CIFAR-10/100 classification accuracy MACE multi-concept setting
Metrics: pre-ASR ASR P4D attack success rate CCE attack success rate FID CLIP score classification accuracy H_c H_a
Baselines: ESD FMN AC UCE SPM SH ED SalUn AdvUnlearn
Builds on: ESD IP-Adapter AdvUnlearn MACE
Closest papers by meaning
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.88)
- Memories of Forgotten Concepts (CVPR 2025 · similarity 0.88)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.88)
- Comprehensive Assessment and Analysis for NSFW Content Erasure in Text-to-Image Diffusion models (NeurIPS 2025 · similarity 0.88)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 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.