AI image generation is getting crazy good

It worked; he’s neatly de-livered.

Paging @Lucas_Jackson

Ha ha, great one. I should have expected you’d come through!

The prompt was

The Count (Sesame Street) suprising Batman (1966) in the Bat Cave. Batman is appalled. Bats are flying around. Iphone 15 photo.

I had to go back and re-look, I didn’t see the bats flying around the first time.

You know what’s strange about that image? Somehow, AI knew to put raised ‘eyebrows’ on Batman’s mask - that don’t actually exist - to emphasize surprise.

I went back just now and changed one thing about the prompt (from 1966 to Burton)

Funny, the thing AI is usually getting wrong is the number of fingers, yet in this case he gave The Count 5… which on the surface seems right - except The Count only has 4! ha ha ha.

The Count (Chocula) suprising Robin (1966) in the Bat Cave. Robin is appalled. Winged bananas are flying around. Iphone 15 photo.

Yes, those eyebrows were on the original mask. The AI might have altered them.

Good catch. My quick glance at batman images didn’t turn it up.

Night Cafe has just added Seedream 4 to their large list of AI models. It is a Pro model there (available only to paying customers, not on free credits) but you get a handful of free samples.

As I’ve mentioned before, Night Cafe also has an option of generating new prompts for you blending elements from your recent prompts and trending prompts from other Night Cafe members, offering maybe a half dozen new options a night. (Some of them interesting, some bland or repetitive.)

I took one of the generated prompts unedited and ran it through Seedream 4 at Ultra resolution and 21:9 aspect ratio

Summary

Photorealism. Medium shot. A young girl with tattered clothes sits on a park bench in a desolate, overgrown urban park. A single, mutated squirrel with glowing eyes watches her from a twisted tree branch. Moody lighting, high detail. Eerie atmosphere, unsettling stillness, digital painting, cinematic shot, muted color palette, volumetric lighting, 8k.

Here’s the result, which clocks in at a beefy 6048 x 2592 pixels.

The same prompt given to Copilot (1536x1024)

And Gemini’s nanner bananer (2816 x 1536)

Here’s a free version of Seedream (with lower resolution and fewer aspect ratios)

Somehow I’m reading “Seedream” as “Sodastream”. That’s not helping. :slight_smile:

Seedream (obviously meant to be “seed ream”) is a product of the Chinese company Bytedance, which also makes Tiktok. (It also also has an AI video generator named Seedance, which is obviously “seed ance”.)

I assumed it was meant as “seed dream”, as in wild images to seed your dreams, or you provide some wild ideas to seed the AI’s dreams that you can watch.

The Sodastream really lifted her mood.

Confession Time: I’ve been following this thread from the beginning but haven’t read hardly a word (I wouldn’t understand) but have been eminently entertained and frequently confused.

Here was Midjourney with the squirrel prompt

Though I liked it more when I ditched the photorealism and went with a more artsy style

Tested Seedream 4 with a couple of my favorite subjects, a possum and the blobfish.

Summary

Realistic iphone 15 photo of a possum attempting to lift a large blobfish standing on a large checkerboard made of squares of sandstone and granite floating on a stormy sea.

It was pretty clear with understanding the basic directions but had problems with the details. It tried to integrate an Iphone into the image instead of just taking it as a style instruction, but it did a pretty good job of an alternating checkerboard pattern. I like the detail of the water droplets on parts of the board lifted from the water. However the possum is mostly raccoon and the blobfish is wrong.

Next is Gemini. The possum is good, the blobfish has problems. The checkerboard pattern is good, but the granite segments don’t look much like granite.

This is an AI called Qwen. The possum is a little ratty and the blobfish is utterly wrong. The checkerboard pattern is good. For some reason everything is out of focus.

Next is Flux. The possum is a lot ratty, the blobfish is weird. The checkerboard pattern is good but the material isn’t.

Here is Copilot and Sora. Great with possums, great with blobfish, pretty good with the sandstone and granite appearances. But it has problems grasping the alternating nature of checkerboard tiles.

And, just for fun, some of the older models. Here is Stable Diffusion XL

The older Stable Diffusion 1.5

And the ancient CLIP + Guided Diffusion, called “Coherent” on Night Cafe

And some of the stone tiles are sponges, for some reason.