Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models
Authors: Die Chen, Zhiwen Li, Mingyuan Fan, Cen Chen, Wenmeng Zhou, Yanhao Wang, Yaliang Li
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experimentation, we demonstrate that our approach achieves superior erasure results with little effect on other concepts while preserving image quality and semantics. |
| Researcher Affiliation | Collaboration | 1School of Data Science and Engineering, East China Normal University 2Alibaba Group EMAIL EMAIL EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1: Growth Inhibitors for Erasure (GIE) Input: A prompt P and a target concept P to be erased. Output: An image xsafe where the concept P has been erased. Encode the prompt as c = Encoder(P) and the target concept as c = Encoder(P ); Draw a sample z T from Gaussian distribution N(0, I); Let [s + 1 : e] be the interval where the token of the target concept is located; w Adapter(zt, c, t = T); for t = T, T 1, . . . , 1 do M DM(zt, c, t); M DM(zt, c , t); I Extract(M , w, s + 1, e 1); Mreplace Inject(M, I); creplace Inject(c, c [s+1:e]); zt 1 DM(zt, creplace, t){M Mreplace}; end Return xsafe z0; |
| Open Source Code | Yes | Our code and data are publicly available at https://github.com/CD22104/ Growth-Inhibitors-for-Erasure. |
| Open Datasets | Yes | In the NSFW content erasure task, we use the inappropriate image prompts (I2P) dataset (Schramowski et al., 2023) to examine the generation results for both implicit and explicit unsafe prompts. ... We also evaluate whether the semantics and quality of the generated images remain unaffected after concept erasure using the COCO-30K prompt dataset (Lin et al., 2014), which consists of 30,000 natural language descriptions of daily scenes. |
| Dataset Splits | No | The paper mentions training an adapter using a limited number of samples ('a few dozen images', '60 prompts') but does not provide specific train/test/validation splits or percentages for these. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments are mentioned in the paper. |
| Software Dependencies | No | The paper mentions using pre-trained models and tools like CLIP, Nude Net, and GPT-4o, but does not provide specific software dependencies with version numbers for their implementation. |
| Experiment Setup | Yes | The training process uses the mean squared error as the loss function, Adam as the optimizer with a learning rate lr = 0.001, and sets the training epochs at 2,000. |