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..

Continuous Concepts Removal in Text-to-image Diffusion Models

Authors: Tingxu Han, Weisong Sun, Yanrong Hu, Chunrong Fang, Yonglong zhang, Shiqing Ma, Tao Zheng, Zhenyu Chen, Zhenting Wang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate the effectiveness of CCRT, we conduct extensive experiments involving the removal of various concepts, algorithmic metrics, and human studies. The results demonstrate that CCRT can effectively remove the targeted concepts from the model in a continuous manner while maintaining the high image generation quality (e.g., text-image alignment).
Researcher Affiliation Academia Tingxu Han Nanjing University EMAIL Weisong Sun Nanyang Technological University EMAIL Yanrong Hu Yangzhou University EMAIL Chunrong Fang Nanjing University EMAIL Yonglong Zhang Yangzhou University EMAIL Shiqing Ma University of Massachusetts at Amherst EMAIL Tao Zheng Nanjing University EMAIL Zhenyu Chen Nanjing University EMAIL Zhenting Wang Rutgers University EMAIL
Pseudocode Yes Algorithm 1 Genetic Algorithm with Fuzzing Input: Initialized Entity Set: E, Optimization Direction: MD, Original and Edited Diffusion Models: ϵθ , ϵθ, Generation Threshold: G Output: Calibration set
Open Source Code No Answer: [No] Justification: We will release our code upon the acceptance of this work.
Open Datasets Yes We use the I2P dataset [38] as the test set to measure the effectiveness of CCRT in continuously removing such content. ... We also extend CCRT to remove three objects (church, tench, and parachute) continuously. ... We utilize Image Net classes because they are public and diverse, not because the algorithm depends on the Image Net hierarchy itself. ... We extend CCRT to remove three randomly selected objects from COCO and CIFAR (church tench parachute) continuously.
Dataset Splits Yes All 2000 images are taken to train the Van Gogh detection classifier, where 0.8 is the training set and 0.2 is the test set.
Hardware Specification No Each removal is a light fine-tune ( 3 GPU-hours).
Software Dependencies No The paper mentions 'GPT-4' and 'Res Net 50' as components/models used, but does not provide specific version numbers for general software dependencies like Python, PyTorch, or CUDA libraries.
Experiment Setup Yes We utilize the Adam optimizer, set learning rate to 1e-4, batch size to 32, and epochs to 30. The final model achieves 90.7% top-1 accuracy.