Text looks great, prompt adherence too, though humans are sometimes horrifically mangled. I don't think this is the same model that the API is using.

The stable diffusion subreddit is throwing a hissy fit over censorship, but that is par the course.

SAI could have easily avooded the massive confusion by providing information on how to use it. Instead, SAI employees are, quite rudely, claiming its a skill issue on discord, but they also don't provide any info on how to use it properly.

If you release some software and users are struggling, telling them to "git gud" and similar is very odd.

Despite all the bad press and turmoil, they keep delivering. Bravo Stability!

I'm not sure that "company which is hemorrhaging money releases yet another open model that won't make them any money" contradicts the financial turmoil, it just means the rate at which they are incinerating money hasn't quite caught up with the rate that investor hype is replenishing their money pile.

I'm not sure if you read the fine article but in fact they do have licensing terms which can provide them income from commercial and enterprise users.

The disdain with which they treat the open-source community means I have no interest in using their commercial/enterprise offerings.

The model is no open, you need to get a license to make money with it.

I wonder if they'll make another release like this. They really seemed to struggle to get SD3 out.

The model looks excellent. Complex arrangements, high quality text, and easier fine-tuning right in time for campaign season... this election year is going to be a fun one.

Let the meme wars begin!

Did anyone figure out what " Limited commercial license" means? FAQ is pretty rushed ("What commercial uses of the Creator License Agreement?")

The images look very detailed, I don't see the typical artifacts in the textures, probably thanks to the better VAE used, I will wait for some anime finetuned model.

What does "medium" mean here?

5 gigabytes vram in its minimum configuration, but various things can be done to increase that. Quantization and distillation might theoretically reduce resource needs, but that's still small enough to get halfway decent CPU generation time.

Is that expected to be superior / on par with SDXL that is much larger?

It's hard to infer relative performance based on parameter count alone. SD3 and SDXL are quite different architecturally. The only way to really tell is to compare it with examples. Even this lobotomized 2B model seems to perform better on prompt adherence and text compared the base SDXL model, so I think it has potential once fine tuned.

2B param model release vs. the 8B param "true" model

They also have small, large and ultra size models. Who knows if those will be made open though.

CivitAI says they've said they will release them. They're apparently not finished yet.