Research map / Image generation / Text-to-image diffusion
Memories of Forgotten Concepts
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.
Shows that concept-ablated text-to-image diffusion models still generate erased concept images from high-likelihood inverted latent seeds, and introduces a relative-distance likelihood metric plus a sequential inversion block to find many distant seeds per image.
Problem
Concept erasure methods for diffusion models are typically evaluated only on output images generated from text prompts with random seeds, so it is unclear whether erased concept information truly disappears from the model's latent space.
Approach
Invert erased-concept images with diffusion inversion (Renoise) to find latent seeds, and measure whether these seeds are likely under the Gaussian prior and reconstruct the image with high PSNR; introduce a relative distance (ratio of EMD of erased vs reference NLL distributions to the standard normal) as a memory metric, and a Sequential Inversion Block to find many distant seeds for one image.
Key results
- Across nine concept-erasure methods and six concepts, inverted erased-concept images reconstruct at high PSNR (PSNR >= 25 dB at dataset level) with likelihoods overlapping reference images.
- The highest relative distance d_N is 2.49 for ESD on Parachute, indicating non-negligible overlap between erased and reference latent likelihoods.
- Sequential Inversion Block yields at least ten distinct seeds per query image with high reconstruction quality, and seed-to-target Euclidean distances have a coefficient of variation of about 2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Invert erased-concept images to retrieve likely latent seeds | Concept-ablated diffusion models · Six concepts: Nudity, Van Gogh, Church, Garbage Truck, Parachute, Tench | Reconstruction PSNR | Helped |
| Use different COCO support images to initialize inversion toward the same query | ESD · Van Gogh | Reconstruction PSNR: PSNRmin = 19.98[dB], PSNRavg = 24.66[dB], PSNRmax = 26.42[dB] | Helped |
| Invert a shuffled church image instead of the original church image | Church-ablated diffusion model · Church | Reconstruction PSNR | Hurt |
| Shuffle church-image patches before inversion and reassembly | Church-ablated diffusion model · Church | CLIP similarity to “church”: 0.3245 -> 0.2906 | Hurt |
| Use sequential inversion to retrieve multiple distant seeds for one erased image | Concept-ablated diffusion models · Six concepts: Nudity, Van Gogh, Church, Garbage Truck, Parachute, Tench | Reconstruction PSNR | Mixed |
Surprising findings
- Some erased models have relative distance below 1, suggesting the erased set is closer to the normal distribution than the reference set.
- Even scrambled church images can be inverted and regenerated, with CLIP similarity to 'church' dropping only about 10% despite a large drop in the concept classifier score.
Limitations
- Analysis assumes white-box access to model weights and the ability to invert the model.
- Multiple-seed search yields lower PSNR and lower relative distance than single-latent retrieval.
- Inversion is powerful enough that it may generalize to out-of-distribution images, complicating interpretation.
Experimental setup
Datasets: I2P COCO
Benchmarks: I2P COCO
Metrics: PSNR relative distance d_N (EMD ratio) NLL likelihood average pairwise cosine distance CLIP score concept detection score
Baselines: ESD FMN Salun Scissorhands SPM UCE AC AdvUnlearn EraseDiff
Builds on: Renoise DDIM inversion Latent Diffusion Models Pham et al. textual inversion analysis of concept erasure
Closest papers by meaning
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.91)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.89)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.89)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.89)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.89)
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.