Enlarging the Capability of Diffusion Models for Inverse Problems by Guidance

The talk introduces two key innovations for enhancing diffusion models in solving inverse problems:

First, a method leveraging two perpendicular 2D diffusion models to tackle 3D problems by representing 3D data as intersecting 2D slices, addressing dimensionality challenges.

Second, a novel solver that uses text prompts to clarify ambiguities, inspired by human perception, dynamically incorporating textual guidance for more accurate solutions.

Experimental results confirm these approaches significantly improve problem-solving accuracy and reduce ambiguities.

Organized by:

baiosphere

Bavarian Academy of Science and Humanities

Helmholtz Munich

LMU Munich

TUM

AI-HUB
LMU

ELLIS Munich Unit

Konrad Zuse School of Excellence in Reliable AI

MCML

Munich Data Science Institute
TUM

Munich Institute of Robotics and Machine Intelligence
TUM

Subscribe to RSS Events feed