Requires COMFY UI
Step 1.
Create or edit your .bat file with the following lines
@echo off
call .\venv\Scripts\deactivate.bat
call .\venv\Scripts\activate.bat
cmd /k python main.py --fp32-text-enc --fp32-vae --bf16-unet
If a FP16 UNET is desired switch --bf16-unet to --FP16-unet
Step 2.
Download FP32 files.
VAE FP32 (SDXL, PONY, illustrious all use this VAE put in the models/VAE folder - Rename if desired)
PONY CLIP FP32 (Put in the models/CLIP Folder)
illustrious CLIP FP32 (Put in the models/CLIP Folder)
Step 3.
Load workflow
Ensure the correct files are selected for the model. In the example the most popular model as of this writing was used cyberrealistic PONY
Step 4.
Save checkpoint with corresponding FP32 CLIP, ensuring that the FP32 launch commands where used. If done correctly your new model will be mixed precision FP32/BF16 or FP32/FP16 depending on your selection.
This is easy to verify as a SDXL base model will be 8.5GB when using BF16/FP32
By Default the saved checkpoint goes in the Comfy/output folder - rename and move it into the models/checkpoints folder
NOTE: The --fp32-vae command is recommended when saving the model however it will cause encoding time increases and may be removed for generation (The FP32 VAE will be downcast to FP16)
Example Images (Seed to Seed)
CyberRealistic v8.5 seed 501
Base seed 501
FP32 CLIP and VAE seed 501
