Fake images

[danielgural]

Quick demo to show how to use embeddings to spot fakes

[jacobmarks]

Should be a separate PR

[jacobmarks]

Upload as raw github user content or google drive and add via URL or download, to avoid bloat in repo. Same goes for all images.

[jacobmarks]

Add the tracking stuff as a separate PR

[jacobmarks]

separate out the tracking stuff

[jacobmarks]

Feedback on the notebook:

• instead of for sample in dataset, use dataset.iter_samples(autosave=True) — it's faster
• We have access to filename as sample.filename. You don't need os.path here
• make sure all images of the app are actually viewable, not blank or pixelating. Add as separate images like the butterflies
• For SD images, maybe use the Text-to-Image Plugin, or make a note about it? This would be more robust but achieve the same effect as your tagging with "diffusion", as well as give transparency into the prompts
• Why installing numba? It should be installed by default, right?
• for compute_visualization(), num_dims=2 is default, "emb" is not a descriptive key, and verbose=True isn't helpful if omitting outputs.

high level feedback:

• some motivation and conclusion would be helpful
• if you are using default model for embeddings, you should say what it is. Also probably worth showing a few models and seeing how things look in the dim reduced space, i.e. what each model picks up on.