10 50
⏳ Running on a free CPU Space — each denoising step takes ~100 s, so 20 steps ≈ 33 min and 50 steps ≈ 83 min. Lower the image-steps slider to finish sooner.
Known memorized prompts (should flag)
Benign prompts (should pass)
How it works. MemGuard reads the diffusion model's score field for your prompt: memorized prompts collapse onto a training image, which inflates the conditional-vs-unconditional guidance direction. The metric combines its norm and a cosine alignment term, calibrated on known memorized / non-memorized SD1.4 prompts.
References
Asthana, R., & Belagiannis, V. (2026). Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability. The Fourteenth International Conference on Learning Representations.