Try using the original adversarial losses for the 2D LDM tutorial

Project-MONAI / GenerativeModels

MONAI Generative Models makes it easy to train, evaluate, and deploy generative models and related applications

Apache License 2.0

621 stars 87 forks source link

Try using the original adversarial losses for the 2D LDM tutorial #268

Open Warvito opened 1 year ago

Warvito commented 1 year ago

In the 3D examples, we had problems with the original LDM's adversarial losses. This does not mean it will not work well for the 2D scenarios.

QuantPrincess commented 1 month ago

@Warvito Can you elaborate more on this? I'm having trouble getting good results for the diffusion part of the pipeline for 3d LDM. My autoencoding model of choice (vqgan) seems to be reasonably converged, but when using the template diffusion model in the 3d LDM tutorial with my autoencoding model, diffusion results aren't quite able to capture the necessary structure and detail of my objects. Did you run into this issue ever? I have an image sice of 64^3.

blofn commented 3 weeks ago

I meet the same problem，3dvqgan’s results are good，but LDM can't generate right latent vector.

QuantPrincess commented 1 week ago

Any updates on this? Can't figure it out!