Attached in both output and JSON form. All you need for this is PixArt-Sigma-XL-2-1024-MS.pth in Comfy checkpoints folder, t5xxl_fp8_e4m3fn.safetensors (from SD3) in Comfy clip folder, and sdxl.vae.safetensors in Comfy vae folder. Also will need the Extra Models For ComfyUI package from Manager installed for some loader nodes. The 6GB VRAM card in question is a GTX 1660 Ti (Nvidia Turing desktop arch, manufactured in 2019), for full clarity here. Have only tested this on the latest ComfyUI with PyTorch rolled back to 2.1.2+cuda121 (in order to make PyTorch Flash Attention actually work on Windows).
#Ridiculously Simplistic Pixart Sigma Workflow (works fine on 16GB system ram + 6GB vram)
heavy nimbus
lunar tangle
U set this up in SwarmUI ComfyUI dlbackend? Or a separately started ComfyUI implementation?