Course Project: CSCI-GA.2271 (Computer Vision)
Institution: New York University
Authors:
- Yuxuan Chen (yc7087@nyu.edu)
- Daniela Pinto Veizaga (dp3766@nyu.edu)
- Amelia Dai (hd2584@nyu.edu)
- De Ren (dr3702@nyu.edu)
This project explores zero-shot harmful content removal in text-to-image generation. Our aim is to detect and remove harmful or sensitive elements from generated images, ensuring safe and controlled outputs without manual curation of harmful prompts.
data/
- classifier/: Contains prediction results obtained from Q16 and NudeNet classifiers.
- images/: Contains generated images from baseline models, controlled methods, and our proposed approach. Folder 'images/inpaint_new_mask_1219_revised_prompt' contains the final results that we report.
- prompts/: Includes the prompt texts used to generate potentially harmful images.
notebooks/
- classifiers.ipynb: Notebook to run and collect classification results using Q16 and NudeNet.
- detection.ipynb: Example notebook for segmenting images and detecting harmful segments.
- images.ipynb: Notebook for generating harmful images.
- prompts.ipynb: Notebook for collecting and managing harmful prompts.
- summary.ipynb: Contains a summary and presentation of experimental results.
- inpainting.ipynb: Notebook for inpainting and visualizing the inpainted results.
- revise prompt.ipynb: Notebook for getting safe revised prompt using LLM in the inpainting step.
- ssim.ipynb: Calculating the semantic similarity.
src/
- Contains executable scripts for harmful segment detection and inpainting.
-
Installation:
- Ensure you have a suitable Python environment set up.
- Install dependencies listed in
requirements.txt.
-
Harmful Content Detection & Removal:
- Use scripts in
src/detection.pyto detect harmful segments in generated images. - Employ the inpainting methods in
src/inpainting.pyto remove these segments seamlessly.
- Use scripts in
For questions or further information, please reach out to the authors via the provided email addresses.