Thanks! about glide-text2im HOT 3 OPEN

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): #2 (comment)

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.

from glide-text2im.

what is it filtered on exactly?

• #2
• #11
• #16

from glide-text2im.

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.

from glide-text2im.