question about 'SD Features'

[develop-productivity]

I'm looking forward to seeing you open up the implementation code for the 'SD Features' model mentioned in the paper. Because I think SD features is a research area.

[Darkbblue]

Hey im trying to implement this by myself but i encounter a problem of understanding the paper. Here's a sentence from appendix: "To extract the intermediate U-Net features, we add a noise equivalent to the 100th timestep noise to the input image and evaluate the corresponding noisy latent using the forward diffusion process. " Does this mean i should simply compute the noisy latent z_t and send it into the unet model? (So the former and latter halves of the sentence seem to be describing the same thing?)

Or should i first add noisy to the original image, then encode it with VAE and treat the result as z_0, and finally add noise again to get z_t?

[develop-productivity]

I think: Firstly, the hidden variable is obtained through VAE encoding, and after a forward process, noise is added to 100 time step (z_t can be obtained at any time through z_0 noise), and then z_100 is denoised in one step to obtain the desired feature.
It is inevitable that something is wrong, looking forward to the original author's discussion!

[Darkbblue]

Hey ive implemented the code at https://github.com/Darkbblue/diffusion-feature. Though it can run on my server, i havent tested the extracted features on image classification tasks. Maybe you want to give it a try?

[develop-productivity]

Thanks. I'll take a look

[Yeez-lee]

Thanks. I'll take a look

Hi, did you do the tests?

[develop-productivity]

Thanks. I'll take a look

Hi, did you do the tests?

No, I didn't do the tests, but I made some modification to fit my project.