Hacker News Reader

Comment by Karawebnetwork | Hacker News Reader

I was curious to compare results with Craiyon.ai

Here is "llama in a jersey dunking a basketball like Michael Jordan, shot from below, tilted frame, 35°, Dutch angle, extreme long shot, high detail, dramatic backlighting, epic, digital art": https://imgur.com/a/7LoAtRx

Here is "Llama in a jersey dunking a basketball like Michael Jordan, screenshots from the Miyazaki anime movie", much worst: https://imgur.com/a/g99G7Bn

View on HN

Craiyon did step up a lot in its understanding recently. The image quality is still not the best but it if you ignore the blurriness, the scary faces, and the weird shapes, it can sometimes be better than dall.e.

Fascinating, are there any other similar products in this same category as DALL.E and Craiyon?

wombo.ai and midjourney

Not products, but Google Research already published papers about two different, better models:

https://parti.research.google/ https://imagen.research.google/

The models themselves are not public however.

Wow. Those models, particularly Imagen, are of an entirely separate calibre. None of that psychedelic foggy memory swirl that is characteristic of the space. I can see why Google Research is hesitant to release them.

After using all of the different models extensively, Stable Diffusion is currently state-of-the-art.

Images are more artistic and less clip art-like than Dall-E, but also don’t have a house style like Midjourney. It’s stunningly good - and open source.

What’s really cool is that the devs have worked hard to optimise the model, so after being trained on 1000 A100s it’ll run happily on an 8gb graphics card or M2 Mac.

Reply on news.ycombinator.com