[BUG] [CorrDiff]: The deterministic sampler scales the conditional input

NVIDIA / modulus

Open-source deep-learning framework for building, training, and fine-tuning deep learning models using state-of-the-art Physics-ML methods

https://developer.nvidia.com/modulus

Apache License 2.0

947 stars 222 forks source link

Closed nbren12 closed 6 months ago

nbren12 commented 6 months ago

I think there is a similar bug to #229 in the sampler:

x_lr should not be scaled by "s". It is not part of the evolved diffusion process, it is a conditional input.

nbren12 commented 6 months ago

cc @mnabian

mnabian commented 6 months ago

Yair made a PR here to fix this: https://github.com/NVIDIA/modulus/pull/373 There were some issues reported by Tao, and we ended up reverting this PR here: https://github.com/NVIDIA/modulus/pull/396

nbren12 commented 6 months ago

@mnabian this is a separate issue that only effects inference.

BTW, I also had a PR on the conditional scaling: https://github.com/NVIDIA/modulus/pull/381