2k Full FP32/FP16 Model+Refiner Comparison
FP32 Refiner in GGUF/Q8 and CLIP-G-Large Pruned
FP16 Image + Workflow
(The only difference is FP16 vs FP32)
FP32 Image + Workflow
(The only difference is FP16 vs FP32)
Difference Visualizations (Grain Extract and Difference)
Analysis
The benefit to the level of detail on full FP32 is clear however the time investment may not be worthwhile to some.
This can be offset running the FP32 model in FP16 mode until you find an image that you would like a increase to level of detail on, then loading that image.
Full FP32 has to be forced in COMFY UI
--normalvram --fp32-vae --fp32-text-enc --force-fp32
For older 8GB cards forcing FP32 may increase generation time considerably but should not be that impactful to more powerful GPU's
With CPU offloading I can run 3 Large FP32 models downcast to FP16 at the same speed or faster in most cases. This is likely do to CPU handling FP32 better and CLIP always being handled by CPU unless forced in GPU