john_noumenonic

7 ms·

1.
by john_noumenonic 3y ago
In the case of pictures taken with your phone, the "true" answer would be "real" as it's not a synthetically generated image, just some post-processing/cleaning-up. The overall classification is going to be aff
2.
by john_noumenonic 3y ago
Wasn't aware of that context in the anime art community, that's good to know about, so thanks. Currently, artist drawn anime images weren't a primary stylistic focus, so the gap makes sense.
3.
by john_noumenonic 3y ago
It shouldn't classify the example you explained as AI generated, but we are looking at expanding some functionality for similar use cases to that. W.r.t AI-upscaling, the current model isn't looking for it specifically, since many
4.
by john_noumenonic 3y ago
Do you know what base model was used to generate the image? Was it SD v1.5?
5.
by john_noumenonic 3y ago
We're generally seeing an accuracy of 90% on test sets, with its distribution primarily being images generated by diffusion models and the "real" images that are more ground level imagery. We'll have to take astrophotogr
6.
by john_noumenonic 3y ago
True enough, though we're interested in stress testing our model to see where the gaps are right now.
7.
by john_noumenonic 3y ago
That's great to know actually, especially the low-quality/distorted images bit.
8.
by john_noumenonic 3y ago
Are you able to use the image upload box to select a file or drag-and-drop images?
9.
by john_noumenonic 3y ago
The preliminary model gets ~93% accuracy on an "in-the-wild" test dataset of ~10k images pulled from the widely used models, i.e. midjourney v5/v6, SDXL, DallE-3, etc.