Do you have a large collection of checkpoints that you ABSOLUTELY CAN NOT DO WITHOUT, but also can't afford another Terabyte of high-speed storage? Why not replace some of your checkpoints with quantized versions?
If you download this, please don't forget to give the original creator some love (like, favorite, donate buzz). I didn't create this (I only quantized it) and I don't want to take attention away from the original creator.
I quantized just the U-net to FP8, VAE and CLIP remain at FP16. In my tests with ComfyUI, this produced output that is essentially identical to the output of the original FP16 checkpoint. If you find settings that lead to output that substantially differs from the FP16 output, please comment those settings. Quantizing CLIP to FP8 works as well, but then the output will definitely change. The quality is similar, but given the small gain in space, I opted for not quantizing CLIP to allow the checkpoint to be a drop-in-replacement for the unquantized one.
The unquantized checkpoint is https://civitai.com/models/137781?modelVersionId=501752
If the original author considers this quantized checkpoint to encroach upon his/her rights, I'll gladly unpublish it on request.