search

Julian-Wyatt / AnoDDPM

CVPR Workshop paper - AnoDDPM: Anomaly Detection with Denoising Diffusion Probabilistic Models using Simplex Noise
https://julianwyatt.co.uk/anoddpm
MIT License
174 stars 28 forks source link

simplex noise #10

Closed CaoYuanpu closed 1 year ago

CaoYuanpu commented 1 year ago

I was wondering why the simplex noise depends on the diffusion timestep t in rand_3d_fixed_T_octaves().

Julian-Wyatt commented 1 year ago

Hi, Yes this was a minor experiment although I’m fairly sure no images generated in the paper use that function. I was considering testing a 3D noise function where the depth of the image is the timestep in the diffusion process, but there wasn’t enough stochasticity to learn from I felt. Hopefully that helps

CaoYuanpu commented 1 year ago

Hi, Thanks for replying. But it looks like you use it in generate_simplex_noise(), which is the default for noise_fn='simplex'. Another question is I was wondering what specific simplex noise you used for MVTEC leather dataset, rand_2d_octaves()?

image
Julian-Wyatt commented 1 year ago

I think in general I used a 3D simplex noise function due to the directional artefacts I mention on line 112. For that there's no massively important depth slice to use so I just kept it as being dependent on t; although in theory, I feel it should work if you pick any random slice if you shuffle the seed each time.