Thanks!

[afiaka87]

Thank you for releasing the filtered model. I realize a lot of people have complained about this, but it's a dicey alignment issue unfortunately. That you did the extra work to release something was very cool.

Feel free to close 👍

[aufbakanleitung]

So what is it filtered on exactly? Because I tried a few things like "Elon Musk as a saint on Mars in the style of Salvador Dalí" but it just created a picture of an Elk with I guess a somewhat Martian color background.

I assume this is because of the filtering. It doesn't know who Elon Musk is. So I'm wondering, what is left? Apart from Corgi's.

[woctezuma]

what is it filtered on exactly?

• 2

• 11

• 16

[Arcitec]

Alright to sum it up:

• The paper's model, "GLIDE", is not released.
• The released model is called "GLIDE (filtered)".
• The paper's model has 10x more parameters (more intelligence/detail). The small public model has 300 million parameters, compared to the unreleased 3.5 billion parameter model.
• The public model can do 256x256. The paper's model can do 256x256 and 512x512.
• The public model has filtered all humans, violent objects (weapons etc), and hate symbols (swastikas etc). It doesn't know anything about what those look like so it can't draw those.
• Here's a comparison between the paper's model (top row of image) and the released model (last row of image): https://github.com/openai/glide-text2im/issues/2#issuecomment-999037054

So the public model can basically make pretty bad output and no humans.

I guess they're never going to release the detailed model since they'll probably turn it into some product they sell to companies later.

Maybe it'll be like GPT-Neo later, where some open source folks take the paper and re-implement the full model and re-train it from scratch on open source data.