November 1, 2024

New Image AI generator looks gorgeous

The new Flux AI image generator has been released from Black Forest Labs and I have to say, it looks absolutely stunning. Previous real life image generating models have always left the final output looking slightly warped or "generated" with misplaced limbs, too many fingers, slightly mis-aligns limbs, but this model looks as close to reality as I've ever seen.

The company claim that the Flux AI strength is its "ability to generate images with exceptional detail and accuracy. From intricate anatomical features to vibrant, true-to-life colors, the platform excels in producing visuals that captivate and inspire. "

The site https://flux1.ai/create is free to register and allows a limited "free" number of images with a wait time of up to 2 minutes at busy periods, but upgrading to Pro will reduce that down to 30 seconds, but on my trial on a free account, I had the image generated in under 15 seconds.

My prompt was:

" An Irish Red Setter sat in a lush garden looking at the camera with a robin red breast sat on its head. In the background there are trees and in one of those trees sits an owl."

An Irish Red Setter sat in a lush garden looking at the camera with a robin red breast sat on its head. In the background there are trees and in one of those trees sits an owl.

The image has a cartoon-like softness and some of the branches in the background, like the one the owl sits on, are non-realistic, the fence panelling in the far bargrounds are stepped or uneven, but these are small details. Overall the image is outstanding and certainly fit for publication.

Anatomy

One of the challenging tasks for AI image generation is with human anatomy, especially around limbs and fingers, so I tried a new approach. This time my prompt was:

"Two people arm wrestling on Plam Beach California. One is a large man, the other is a slight lady. She clearly has the more strength and you see the worry on the male wrestlers face as she looks like she is going to beat him. A crowd looks on shouting encouragement at the girl"

Two people arm wrestling on Plam Beach California.

Overall, the image was impressive. The skin was softer than I would have liked and the "wrestlers" thumb is in an awkward position. The skin tones looked good and even the whiter casts under the wrestlers arm was very realistic with good defined anatomical muscle definition. The girls elbow was well pronounced with good shading. What was a little disappointing was the complete ignorance of my prompt asking for "A crowd looks on shouting encouragement at the girl" but given the tight crop of the image, maybe it boxed itself into a corner. And the obvious issue is that the whole Arm wrestle concept was not understood. It would have been good in a fist-bump competition though!

I ran the prompt again, this time with a wider aspect of 16:9, this time it generated a wholly different image with poorer anatomical detail. In this one, you can see the inaccuracies of the hands and how they interact with each other. This has always been the Achilles heel of AI generated images, and while Flu1 has improved it significantly, there is clearly still some work to be done on this front.

Two people arm wrestling on Plam Beach California.

Text

Finally we move onto the real problem we have always seen with image generators and that is with text. A seemingly impossible job is rendering text in a legible manner. Anyone that uses Dell-E or the Image Generator within Co-Pilot would attest to. Let's try it out on a topical image.

Prompt: Two beautiful girls smiling at the camera, one wearing a red cap that says "MAGA" on it, the other girl wearing a blue cap that says "Kamala"

Two beautiful girls smiling at the camera, one wearing a red cap that says "MAGA" on it, the other girl wearing a blue cap that says "Kamala"

Things of note here: The text has come out perfectly, if not the cap. The girls have come out very caucasian. Compare this to Dall-E output - we cannot generate the same image as Microsofts policies prevents the use of political stateents on it's generated images, so we have swapped out MAGA and KAMALA as Blue and Red.

It struggles more with the text, but you see how cartoon looking the girls are. But! Their ethnicity is certainly more mixed than Flux1.

Two girls wearing a red and blue hat

Conclusion

So overall, the Flux1 image generator is very impressive. It's still lacking on some of the inhibiting aspects of AI image generation like interpretation and the rendering of true-life anatomy, but it's another step closer to image perfection.

You can register a free account over at https://flux1.ai/create or register for a pro account that will cost $15.9/ Month and give you 800 generations per month.

Do yiou have a favourite image generator? Which one do you use/recommend and for what use?


Leave a Reply

Your email address will not be published. Required fields are marked *

Are you tech-ready?

Microsoft Certified Partner

Technologies

Microsoft 365
Exchange Online
Sharepoint
Teams
Defender
Workflow integration with popular office applications. Our goal is total synergy to remove operational bottlenecks and have the environment as clutter-free as possible.
+44 794 7110 612
envelopephone