FaceID workflow question
Hello.
Any particular reason why you're using the INT8 quantized version of the v1.4 10Eros model instead of the full fp8 checkpoint? In the new FaceID workflow I mean.
I would hazard a guess that it's because the community has been clamoring for an INT8 ConvRot version of Eros for a while now (myself included) and he delivered us one. It's just a superior format for quality and efficiency.
It's much faster and better quality than fp8.
Is this only noticeable on the FaceID workflow?
I'm asking because I know the theory, but in practice I've been comparing the results using the regular v5 DMD workflow with the fp8 model, versus using the int8, and while int8 is indeed (slightly) faster, the results are quite lower quality overall for the same exact setups and seeds. Tried both with the same vaes and text encoder as the newer FaceID v2 workflow, and also with the extracted vaes and encoder from the older full fp8 v1.4 checkpoint, and got the same results in either case. FP8 is consistently giving better results than the INT8, even as the theory behind the int8 convrot quantizations says otherwise.
Maybe I'm doing something wrong? are some weights or the number of steps or the sampler expected to be different for the INT8 model?
I got a lot of limb distortion from fp8 that I don't get on int8, it could just be a luck and scene specific, bf16 was the actual overall better quality result but slower.