I dont understand how this article is presenting this new model as a competitor of SD. SD already has LCM support, there is StreamDiffusion which you can run on your computer to see near live img2img. There is Control Net, Turbo and lots of other things community has been doing on SD to improve performance.

This article is very thin on details of PixArt and it only made me think this is just a fork/variation of SD instead of a new thing like SD.

Yes, it's mostly a new training technique (that is impressive: "PIXART-α only takes 10.8% of Stable Diffusion v1.5's training time"). I'm not really sure if it really improves the image quality over SDXL, but it may a bit: https://pixart-alpha.github.io

From the comparisons, this seems quite a bit better than SDXL-LCM. I wonder how long it will take projects like Fooocus to add support for it, I'd be very interested in trying it out.

You can use it in comfyui. Instead of the checkpoint loader use the unetloader. The model has a different unet architecture than standard sdxl models.

On how many images was this one trained? I remember their last model being quite undertrained compared to stable diffusion models

It's Latent Consistency Model (LCM) and ControlNet added to PIXART Alpha.

https://arxiv.org/abs/2401.05252

I was kind of disappointed to dive into comfyui and try to make some game assets with sdxl. ControlNet is a huge leap. But overall outputs aren’t quite there yet.

I was unable to verify if this model is trained only with public domain images? Anyone found this information?

Wonder how long it'll be before we start getting 3D scene generators (with the 3D models), that work as well as these image generators?

Is it just me, or do 1 step comparisons not really help? I need at least 10 steps to get anything useful out of SD or SDXL?

It depends of you're using a model distilled for fast inference like LCM, LCM-LoRA or turbo.