What You May Have Missed #7

DALL·E API & Midjourney v4 / 3 new Gen AI apps /AI ethicists' hardships / When generative AI becomes unethical / WebSummit: Gary Marcus and Noam Chomsky / Meta protein folding model / Google AI@ '22

“Robots telling stories at night around a campfire.” Credit: Author via Midjourney v4

DALL·E API, Midjourney v4, and the benefits of hiding prompts

OpenAI finally made DALL·E available through an API. This news comes quite late given the popularity of Stable Diffusion (SD), but it’ll still spark the emergence of new gen AI companies. The reason is DALL·E—in contrast to SD—removes the burden of “good prompting” from the user by hiding additions they automatically include to make the images more appealing.

Levelsio (creator of InteriorAI and AvatarAI) tweeted about this recently: “most of us already automated prompt writing away with a front end interface with big buttons and selectors. Regular people don't have the time to figure out prompts.” I agree that, although prompt engineering will be ubiquitous, the ability required to obtain good results will go down over time, as companies hide the complexity of prompts behind the scenes.

@levelsio @levelsio

There's a reason Dall-E auto converts "a dog" In background to: "4k hd, high detail photograph, shot with sigma f/ 4.2, 250 mm sharp lens, shallow depth of field, subject= white golden retriever, consistent, high detailed light refraction, high level texture render"

Midjourney does something very similar to always generate beautiful images—and they just took it to the next level with the release of v4. The new version is significantly better than anything I’ve seen. Here’s a side-to-side comparison of “a penguin in Venice” between Midjourney v3 and v4:

Lynn Cherny @arnicas

Holy crap, the leap in new model v4 on @midjourney. This is thumbnails of a penguin in Venice, v3 on left, v4 on right --no specific style request. #AIart

Midjourney v4 was trained from scratch (it doesn’t use SD or the DALL·E API).

Midjourney @midjourney

@WeirdStableAI V4 is a new model trained from scratch by Midjourney. Totally new codebase. Been in the works for 9 months

It can handle more complex prompts, it’s better with small details, and, maybe most importantly, it’s better with multi-object scenes:

Stephen Young @KyrickYoung

Midjourney v4 can handle multiple subjects in the scene with a high degree of coherency 🤯😮 "A gnome and a robot playing chess in the park" #midjourney @midjourney #aiart #MachineLearning

Although it doesn’t seem to have mastered compositionality:

“A blue cube on top of a red cube.” Credit: Author via Midjourney v4