A small selection of quantized SDXL models in GGUF format, to be used with the custom ComfyUI nodes from https://github.com/city96/ComfyUI-GGUF
| Model | Hugging Face | CivitAI |
|---|---|---|
| CyberRealistic XL v4 | hf link | civit link |
| Juggernaut v11 | hf link | civit link |
| RealVisXL 4.0 | hf link | civit link |
| RealVisXL 5.0 | hf link | civit link |
| SDXL 1.0 | hf link | |
| ZavyChroma 5.0 | hf link | civit link |
| ZavyChroma 10.0 | hf link | civit link |
These are made with llama-quantize from llama.cpp b3600 patched with https://github.com/city96/ComfyUI-GGUF/blob/main/tools/lcpp.patch
A couple of models are quantized with llama.cpp b3962 patched with lcpp_sd3.patch instead, which may give slightly different results than the default b3600-quantized ggufs but not necessarily better. These are marked in the filenames; most users won't need these.
It seems that Q4_0 often gives good results for SDXL models, usually coherent but not always consistent with ones from an unquantized model. Adding Q5_K_M because it sometimes gives results clearly closer to the original F16 than Q4.
Repo includes basic clip-l and clip-g, and also an improved SDXL VAE called xlVAEC_c91, found here: https://civitai.com/models/152040?modelVersionId=360299 which can be recommended as it seems to give better colors, contrast and details compared to base.
If model creators or others wish to have any GGUF files added or removed, I will consider such requests when I see them.
There are also quantized models uploaded by another user at https://huggingface.co/Old-Fisherman/SDXL_Finetune_GGUF_Files