Misalignment between pesudo code and implementation

NVlabs / edm2

Analyzing and Improving the Training Dynamics of Diffusion Models (EDM2)

Other

466 stars 18 forks source link

Misalignment between pesudo code and implementation #5

Open WANGSSSSSSS opened 3 months ago

WANGSSSSSSS commented 3 months ago

Hello, Karras, congratulations on being chosen as the CVPR best paper candidate for EDM2. But is this a bug? As there is no in-place ops, so w in these two figures are different.

Al-Kharba commented 2 months ago

w is not a copy of the original weights, it's pointing to the same weights. See the simple example below.