Nanosaur is not a semantic model
Despite the links to the research papers and misleading module names, the repository has only succeeded in creating a larger, overweight module that performs the same function.
In the second part of the VAE training, the high-frequency image information bleeds into the latent.
The resolution independent semantic information is destroyed for the sake of accurate image output.
In short, the model description does not fulfill its advertised purpose. The training code has an extra step that obscures the process without improving it.
The code is faulty, feature creep, both the latent and the generated images have the same shortcomings as previous models.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for readyforbattle/nanosaur-has-no-semantic-autoencoder
Base model
well9472/Nanosaur-250M