Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Exploiting the Signal-Leak Bias in Diffusion Models
Learn how a signal‑leak bias causes mismatches in Stable Diffusion, and discover a simple inference‑only fix that improves prompt fidelity without retraining, with a live demo.
I will show a Jupyter Notebook demo of my research work on a bias present in most diffusion models for image generation, especially Stable Diffusion (v1 and v2).
I will show that there is currently a discrepancy between training and inference processes in these models, and show how to fix and exploit this discrepancy to gain more control over generated images. The method does not require any additional training and can be applied directly during inference.
Project page: https://ivrl.github.io/signal-leak-bias/
Research paper: https://arxiv.org/abs/2309.15842
Github: https://github.com/IVRL/signal-leak-bias
Exploits diffusion model's signal-leak bias for training-free style and diversity.
Exploits diffusion model signal-leak bias via initial latent manipulation, improving generation.
Compose Email
Loading recent emails...