How Stable is Stable Diffusion under Recursive InPainting (RIP)?🧟

Javier Conde1         Miguel González1         Gonzalo Martínez2         Fernando Moral3         Elena Merino-Gómez4         Pedro Reviriego1
1ETSI de Telecomunicación, Universidad Politécnica de Madrid, 2Universidad Carlos III de Madrid, 3Universidad Antonio de Nebrija, 4Universidad de Valladolid
Paper Zenodo

🌟 Official Demo: GenAI Evaluation KDD2024 🌟

Welcome to our official demo for our research paper presented at the KDD conference workshop on Evaluation and Trustworthiness of Generative AI Models.

This demo shows the effects of recursively applying inpainting with a random mask to an image. A mask is applied at each iteration to remove a random part of the image and subsequently, inpainting is used to reconstruct the image. As iterations progress, the image can change significantly. You can see the effects of two iterations on "The Nobleman with his Hand on his Chest" by El Greco. Now is your turn, play with images, mask sizes, and iterations to see the effects of recursive inpainting!

🚀 How to Use

  1. 📤 Upload an image or choose from our examples from the WikiArt dataset used in our paper.
  2. 🎭 Select the mask size for your image.
  3. 🔄 Choose the number of iterations (more iterations = longer processing time).
  4. 🖱️ Click "Submit" and wait for the results!

📊 Results

You'll see the resulting images in the gallery on the right, along with the LPIPS (Learned Perceptual Image Patch Similarity) metric results for each image.
2 5
Mask Size
LPIPS Distance
Examples

🎨✨To cite our work

@article{conde2025recursive,
   title={Recursive InPainting (RIP): how much information is lost under recursive inferences?},
   author={Conde, Javier and Gonzalez, Miguel and Mart{'\i}nez, Gonzalo and Moral, Fernando and Merino-Gomez, Elena and Reviriego, Pedro},
   journal={AI \& SOCIETY},
   pages={1--17},
   year={2025},
   publisher={Springer}
}