re: Bing's Image Creator is Woke! NSFW

This board needs more of this quality content

Did you try replacing the term ’ray bands’ with ‘aviators? May be a branding issue.

You should get the real deal ….

I seem to have found a hole in the woke that I expect to be filled shortly.

Dirty old man

quote:

---

If you tell it what to say. But, as in the op, unprompted text is just gibberish

---

That’s mostly true, yeah. Even the best models today will spit out a lot of nonsense for unprompted text, especially when theres a lot of it (like sponsor names on a racing suit).

But it wasn’t that long ago that the models struggled to do text at all even when specifically prompted.

Baw, unless you orders that, she ain’t got no teef!

The porn "stars" better get to striking to keep AI out of their movies or they all are going to be back to street-walking.

quote:

---

I'm working on a new novel

---

What do you write? Fiction/Non-Fiction, what genre?
Would you share any titles? I know that kind of means outing yourself on TD so understand if you don’t.

quote:

---

I wonder if the weird text is just copyright stuff - they don't want to use any real Trademarked name? It just seems too bad to be due to lack of ability?

It can get all that detail, but not spell "Renault"?

---

Nah, it’s just a weak point of diffusion models in general. I don’t really want to go into the whole spiel about how they work right this second, but it basically comes down to this: by default, diffusion models are kind of bad at text, and particularly unprompted text.

This is partially due to the overall architecture of the models, and partially due to the training process. Imagine someone is training a diffusion model with NASCAR photos. The photos all have a bunch of different sponsor logos, with similar layouts/fonts but different words. It’s exceedingly unlikely that the training captions will describe the text of every single logo. So as a result, you wind up with a model that knows race cars have logos but has no idea what the logos actually say. You wind up with a pretty incoherent average approximation of what the model thinks a generic logo should look like.

The architecture issue has largely been solved by today’s state of the art models. There are several models that are really good at generating text when you prompt for it. But the issue of unprompted text being jumbled still exists because of the training process, along with technical limitations on the number of parameters a model can realistically be trained on.