Hacker News Reader

Comment by amelius | Hacker News Reader

But why pick a dog as an example?

Humans are much worse in telling dogs apart than other humans (except perhaps the owner of the particular dog).

So for all we know, the AI didn't generate a portrait of this particular dog but instead a generic picture of this breed of dog.

View on HN

Mostly because I thought of it more as an art project than a technical accuracy project. However, the honest answer to your question, is because I have a ridiculous amount of photos of my dog on my phone . Getting training data is hard work.

But this is totally true, I found that maybe 30% of the images I generated did not look like my dog at all. However the rest do a good job at capturing his eyes and facial expressions that he actually makes. I thought that the chosen image I worked from captured the look of his eyes super well.

But yeah, nobody but me would really appreciate that.

Because you invent a new word when you train dreambooth and teach it that your subject is an example of that word. The fact that the word you've created returns photos similar to subject is a sign that it worked.

I suppose that dreambooth is pretrained on a large dataset that includes many different dogs.

My point is that it is difficult to judge (for us) that the returned photos are actually similar to the subject.

The paper shows dogs with very distinctive fur coloring. Particularly the corgi with a white strip between its eyes. I think the paper would be completely fraudulent if this dog were also featured heavily in the training set. So the point is the white stripe corgi isn’t in the set, and with a few examples, the model could then generate brand new images of corgis with a similar fur pattern. Maybe all it can do is fur patterns but it’s a start.

I linked this elsewhere but here are Pokémon image generations of my (mutt) dog: https://imgur.com/a/11OxoSA

She’s pretty unique looking and it comes through even with heavy styling.

Reply on news.ycombinator.com