Stable Diffusion 3 Medium is here, and it’s BAD

Stable Diffusion 3 Medium is here, and it’s BAD

On June 12, 2024, Stable Diffusion 3 Medium, the latest version of the well-known AI image generation model, trained with 2 billion images, was launched. After months of anticipation and promises, we decided to try it out, and here are some images of what we managed to produce.

The announcement email triumphantly states, “Overall Quality and Photorealism: Overcomes common artifacts in hands and faces, delivering high-quality images without the need for a complex workflow”, and “Typography: Achieves unprecedented text quality with fewer errors in spelling, kerning, letter forming, and spacing by leveraging our Diffusion Transformer architecture.”

We immediately put it to the test, and the results were “surprising.”

First, we wanted to test the quality of the text. The email included this image along with the prompt, which we immediately replicated using the exact same words.

It looks exceptionally well done, but how is the reality of things? This is what we obtained:

This goes beyond the worst expectations. Disappointed by the text, we wanted to test their “perfect hands and faces,” and here is where the real fun begins.
Let’s see how SD3 can do hands:

We let out a loud, bitter laugh. But maybe, MAYBE, it can do better with people. Can it? Well… how about a handshake between businesspeople in an office?

This is NICE! The guy has three arms, while the woman has a left right-arm. Also, the arm itself isn’t right, and sorry for the play on words. Both have enough fingers to sell them.

This article is getting a little too harsh, so we need some love. That’s where the “Smiling african female young doctor is making a heart shape with her hands in a medical office” prompt comes into play!

Man, this is gross. This is not a human, this is a crab. And a creepy one too. But hey, this is only the beginning! Let’s try something else, a crowd of teenagers outdoor with arms raised. At this point, we already know what to expect…

Please look at those facial expressions. They are… beautiful. But making crowds is difficult, so let’s stick to a single person: “Angry teenager screaming”.

Here you can see the “Overall Quality and Photorealism: Overcomes common artifacts in hands and faces, delivering high-quality images without the need for a complex workflow” feature at its best.
We don’t surrender. We want a businessman smiling in his office. No mention about hands at all. We got this:

This is a severe genetic malformation that causes a man to have his hands fused into a single limb. He definetely need cures. But I’d stay away from the doctor above.
Let’s try some sports! We are Italian, so we love soccer. “Soccer player ready to kick a free kick”.

We doubt that this guy can even manage to kick a ball. The ball itself is way bigger than it should be, but it’s really not the main problem.
Jeez, we’re getting tired. Let’s get some relaxing images. It’s summet now, so we imagined a woman bathing in the sun on the beach. And SD3 immediately give us its result:

Imagine approaching this woman to flirt. Everything is so wonderful.
By the way, you can have your relaxing time even if it rains, or it’s cold. There are spas and saunas! What about a man enjoying a sauna bath in a luxury spa? After all, he deserves it after long and tiring workdays. What he does NOT deserve, however, is his anatomy. This is by far the most beautiful image we’ve ever obtained.

It’s… the best thing we’ve ever seen. Really.

Stable Diffusion 3 is a disappointment, an insult to everyone who waited months for this product. A completely unusable engine. By far the worst product in recent years, vastly inferior even to its early editions. A collection of deformed limbs, severed, disfigured faces that wouldn’t look out of place on a final boss in Splatterhouse, a 90s horror arcade videogame.

Use any other model and don’t waste your time on this piece of crap, unless you want to laugh a bit.