ChatGPT and DALL-E 3 [video]

I played with this. I can see the place for it, but I like my own prompts better than the ones it makes. IIRC you could tell it to knock off the prompt composing and just send through your exact prompt with some limited success. eg generate an image using the dalle plugin of a bottle of water

    {
      "n": 1,
      "prompts": [
        "A transparent plastic bottle of water, filled to the brim, with a blue cap. The label on the bottle is white with blue lettering. The bottle is set against a neutral background, and the light is reflecting off the water inside, giving it a fresh and clean look."
      ]
    }

Honestly, when I first heard of this integration, I'd thought that this was going to be a new way of controlling generated images - like if you could say "move the subject a little to the left" or similar - the kind of stuff you currently need ControlNet for. Instead, this is more of an automated "prompt composer" - still interesting, but I hope that they explore the idea of letting users have more control over the outputs sometime.

Interesting.

This is like DeepFloyd but probably combined with OpenAI's strength in NLP field.

I remember Ilya Sutskever often mentioned how multimodal is important in multiple interviews. ControlNet can produce more impressive results for sure, but the model being able to have strong understanding of multimodals like language and light and space as a unified modal will push the industry forward towards the goal of AGI.

“That’s great!

The kids are asleep and bears have taken my family hostage unless Larry is BBQ’d. What recipe should we use? What does the spread look like?”

When do experts predict these will be able to generate 5 minute long video?

Is this many years out due to computation cost?