this post was submitted on 31 Oct 2023
1 points (100.0% liked)

Machine Learning

1 readers
1 users here now

Community Rules:

founded 1 year ago
MODERATORS
 

There's a lot of image generators that'll allow you to generate multiple images based on the input data. Is there something that'll generate an image that's an average of the trained model instead?

Didn't plan on using machine learning for this project but realized it might be an interesting path to explore.

top 6 comments
sorted by: hot top controversial new old
[–] bob2013sherland@alien.top 1 points 1 year ago

Maybe train a Generational adversarial network (GAN) with your own data, that way you can control what it learns / generates

[–] nikgeo25@alien.top 1 points 1 year ago (1 children)

If the model has the following format:

Encoder, diffusion in latent space, decoder

you can average the embedding right before the decoder. Just make sure your averaging is of a representation that can be meaningfully averaged. E.g. averaging images would result in garbage, but their corresponding embeddings might provide something useful.

[–] max_b_jo@alien.top 1 points 1 year ago

Yup, it's pretty possible that the generated image will result in pure garbage. But it's worth exploring and might be a nice visualization!

[–] PM_ME_YOUR_BAYES@alien.top 1 points 1 year ago

Why not averaging the aligned data then?

[–] muxamilian@alien.top 1 points 1 year ago

My idea was to train a GAN and sample the center point: Just input a vector of all 0s into the generator after training: https://muxamilian.github.io/Robo99/

[–] fulowa@alien.top 1 points 1 year ago

this will just produce garbage no?