illustriousXL (12GB FP32)
This is a full checkpoint using all Floating Point 32 models - Do not load additional VAE or CLIP
If your looking for the CLIP only it can be sourced from this model or here.
UNET can be safely downcast to BF16 or FP16, but using a FP32 CLIP command is recommended especially for COMFY users as it cost no generation time
Description
FAQ
Comments (6)
I'm not entirely understanding of the technical side of things here. Compared to the illustrious that's less in gb would this then be basically the same thing but more/finer details cause it's a full size? Also curious, would it be better now to train on this for lora, would there be improvement as compared to using downcast version?
Downcast UNET will be a slightly different model then base illustrious, It would train better lora's with FP32 in most cases but may not be worth the time investment for some. The GGUF model also would be have more detail.
What are the difference with base illustrious?
The FP32 CLIP is the biggest change