Dive into the details and code:
📖 lnkd.in/d4FjwFY7 (paper)
🖨️ lnkd.in/dkdabrZ9 (pre-print)
💾 lnkd.in/dfnUk8-C (code)

#FAU #DFG #MachineLearning #ComputerVision

Visual results

Results: Extensive evaluation on ImageNet and CIFAR-10 datasets shows superior performance in filtering low-quality samples and improving generation quality.

📄 Paper: lnkd.in/d7JBSkiz
💻 Code: lnkd.in/dqFbC2nP

Online demo coming soon!
#FAU #MachineLearning #ComputerVision #DiffusionModels #WACV2025

💡 Theoretical Insights: We show that our uncertainty estimates are related to the second-order derivative of the diffusion noise distribution, providing a solid mathematical foundation.

🎯 Guided sampling: We use uncertainty estimates to drive the sampling process towards higher quality generations, resulting in improved FID results and fewer artefacts.

Key highlights:

📊 Training-free uncertainty: Our method estimates pixel-wise aleatoric uncertainty during the sampling phase without requiring any model modifications or additional training, allowing to filter out low quality samples.