I haven't read much about how these systems work yet, so this is probably a novice question, but I'd be interested to hear more about how the algorithm handles text input and feeds it into the generator. Does the training process include a ton of tagged images or something, and then the model learns to be able to generate stuff that corresponds reasonably to those tags?

How that is different with DALLE ? https://openai.com/blog/dall-e/

in terms of ai, this uses latent diffusion models: smaller model, does not use clip, and it was trained on noisy data.

Why invent your own models when DALL-E exists? Is it supposed to be better/more-accurate?

We are using already existing models based on open-source research. We have other models that OpenAI does not offer (e.g. super resolution) but they’re not in the app.

The reality though is that DALL-E 2 is still not open for access; open-source models are.

If you know how to get API access—besides waitlist of course—please let me know!

What image dataset/generation library does this use?

latent difussion models (see: https://github.com/CompVis/latent-diffusion).

we also use our own open-source library https://github.com/thegeniverse/geniverse

The image wouldn't generate for me

had issues with the queuing system... fixed!

This is awesome!