Research map / Image generation / Text-to-image diffusion

EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers

ICML 2025 · Image generation › Text-to-image diffusion · approach: Concept erasure in diffusion models · 0 citations (OpenAlex)

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 EraseAnything, a bi-level LoRA-based concept erasure method for flow-matching transformer T2I models (Flux.1 [dev]), combining ESD-style erasure, attention-map suppression, and a reverse self-contrastive loss over LLM-proposed irrelevant concepts.

Problem

Concept erasure in T2I models is hard to transfer from SD v1.x U-Net diffusion models to flow-matching transformer models like Flux, where no explicit cross-attention exists and T5 word embeddings are poor for word-level similarity; erasure must also preserve unrelated concepts.

Approach

Formulate erasure as bi-level optimization: a lower level erases target concepts via ESD loss plus attention-index suppression on LoRA weights, and an upper level preserves irrelevant concepts using a reverse self-contrastive loss over synonym and LLM-sampled irrelevant-concept attention features, with shuffled prompts to avoid index overfitting.

Key results

• On I2P nudity, EraseAnything yields 199 detected nudity images vs 605 for original Flux.1 [dev] and 173 for UCE (second lowest), with FID 21.75 and CLIP 30.24 on MS-COCO 10K (better than original 21.32/30.87 only on FID-relative terms; CLIP slightly lower).
• On entity/abstraction/relationship erasure, Ours reduces Acc_e to 12.5/21.1/18.4% vs CA's 14.8/25.2/22.7%, while raising Acc_ir to 91.7/90.5/90.2%.
• Celebrity ablation: full loss achieves Acc_e 14.9% and Acc_ir 88.5%, the best trade-off among tested loss combinations.

Measured effects reported in the paper

Surprising findings

• T5 word-level embeddings give implausible nearest neighbours for 'nude' (e.g., 'lean', 'deer', 'girl'), unlike LLM-proposed synonyms.
• Flux.1 [dev] is largely insensitive to word order in prompts, and deterministic attention-map column zeroing is easily circumvented by misspellings, affixes, and repeated concept words.

Limitations

• Combining many LoRAs (10+) with normalized sum weakens each concept's erasure effect.
• No fine-grained control over erasure strength.

Experimental setup

Datasets: I2P MS-COCO 2017 val (10K captions) CelebA

Benchmarks: I2P MS-COCO 10K CelebA-based celebrity erasure

Metrics: NudeNet detections FID CLIP score classification accuracy Acc_e Acc_ir Acc_g user study scores

Baselines: CA ESD UCE MACE EAP Meta-Unlearning Real-Era

Builds on: ESD EAP UCE Flux.1 [dev] LoRA

Closest papers by meaning

• Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models (NeurIPS 2026 · similarity 0.91)
• EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.91)
• Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.91)
• Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.90)
• All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.90)
• Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.90)