Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models
Authors: Shristi Das Biswas, Arani Roy, Kaushik Roy
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present the results of our method for erasing artistic styles, objects, identities, inappropriate concepts as well as resistance to red-teaming attacks, as illustrated in Fig. 1. We use Stable Diffusion-v1.4 (SD-v1.4) (3) as our primary T2I backbone, following recent work (34; 35), and set α = 2 for all experiments expect NSFW concepts, where α is set to 5 for stronger erasure. Artist erasure To evaluate the efficacy of style unlearning for mitigating artistic imitation and potential copyright violations, we follow prior works (34; 35) to use 20 prompts each for five classical artists (Van Gogh, Pablo Picasso, Rembrandt, Andy Warhol, and Caravaggio) and five modern artists (Kelly Mc Kernan, Thomas Kinkade, Tyler Edlin, Kilian Eng, and the series Ajin: Demi Human), all previously reported to be mimicked by SD (12). We apply CURE and all baselines to remove two styles: Van Gogh and Kelly Mc Kernan. Evaluation uses LPIPS scores (68), reporting LPIPSe (on erased artists) and LPIPSu (on unerased artists), where a higher LPIPSe indicates stronger removal of the target style, and a lower LPIPSu reflects better preservation of unrelated artists. Following (23), we additionally use GPT-4o to classify artistic styles of the generated images. Acce shows how often the unlearned style is still predicted lower is better. Accu measures accuracy on non-erased styles higher is better. As seen in Tab. 1 and Fig. 4(a), CURE achieves effective target erasure with minimal impact on unintended styles as well as impressive specificity in preserving normal content of COCO-30k (69), outperforming baselines. |
| Researcher Affiliation | Academia | Shristi Das Biswas*, Arani Roy*, Kaushik Roy Purdue University EMAIL |
| Pseudocode | No | The paper describes mathematical formulations and algorithmic steps but does not present them in a clearly labeled 'Pseudocode' or 'Algorithm' block or figure. |
| Open Source Code | Yes | Project Page at https://sites.google.com/view/cure-unlearning/home. ... The code is already attached to the supplementary material for access to reviewers at present, and the paper also provides detailed descriptions of the datasets used and evaluation protocols, allowing others to replicate the setup. The experimental methodology, unlearning targets, and hyperparameters are fully disclosed. Full open access to code and pretrained models, with reproduction instructions, is planned upon acceptance to ensure faithful reproduction by the community. |
| Open Datasets | Yes | Benchmarking against prior approaches, CURE achieves a more efficient and thorough removal for targeted artistic styles, objects, identities, or explicit content, with minor damage to original generation ability and demonstrates enhanced robustness against red-teaming. ... I2P dataset (16) ... Imagenette classes (71), a subset of Imagenet classes (72). ... COCO-30k (69). |
| Dataset Splits | No | The paper uses established datasets like I2P and COCO-30k and mentions generating a certain number of images per class (e.g., '500 images per class') for evaluation, but it does not explicitly define training, validation, or test splits for these datasets within the main text, often relying on existing dataset structures or inference-only scenarios without explicit splits. |
| Hardware Specification | Yes | Evaluated on an A40 GPU for 100 iterations. ... We use standard licenses from the community and provide the following links to the licenses for the datasets and models that we used in this paper. ... As shown in Table 5, the entire CURE operation, including all steps above, completes in under 2 seconds on a single GPU. |
| Software Dependencies | No | The paper mentions using Stable Diffusion-v1.4 as the T2I backbone and tools like GPT-4o and Nude Net for evaluation, but it does not specify any software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | We use Stable Diffusion-v1.4 (SD-v1.4) (3) as our primary T2I backbone, following recent work (34; 35), and set α = 2 for all experiments expect NSFW concepts, where α is set to 5 for stronger erasure. ... For each target object (e.g., French Horn ) is treated as a forget concept cf, without any additional retain concepts cr. To measure the effect of erasure on both the targeted and untargeted classes, we generate 500 images per class and evaluate top-1 accuracy using a pretrained Res Net-50 classifier (73). |