Update - Removed the Garbage Collector nodes following advice in the comments. Also added a Face Detailer after the second refiner step, a Save node in case you want to do post detailing work without it being upcaled and a Preview Selector node to choose which (if any), images to pass to SUPIR.

--------------------------------------------------------------------------------------------------------------

This workflow was the base that I built upon -  https://civitai.com/articles/5355/installing-comfyui-for-use-with-pixart-sigma  Instructions for installing and getting PixArt set up properly can be found on that article.

I'm using the PixArt-Sigma-XL-2-1024-MS checkpoint model from here:  https://huggingface.co/PixArt-alpha/PixArt-Sigma/tree/main

For T5, create a T5 folder in models and get models.safetensors and config.json from here and put them in that folder:  https://huggingface.co/city96/t5-v1_1-xxl-encoder-bf16/tree/main

Get the VAE here:  https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/sd-vae-ft-ema  

The workflow worked nicely but wanted to see if I could improve it by adding SDXL refiners and an upscaler.

It starts with the Pixart generation (note: I have the T5v1.1 Loader node set to GPU, if you have a lower VRAM card change this to CPU). It then goes to a SDXL settings group where you can set the checkpoint model and LoRAs. If you do not want to use a Hyper 8 step LoRA, increase the steps in the refiner Ksamplers.

It gets a small 1.1 latent upscale and is then passed through a step refine and then onto SUPIR for Upscaling. I have the SUPIR set to HighVRAM, drop this unless you have enough VRAM.

Really happy with the results I'm getting on this.