CFG > 1 generates error "output with shape [1, 8, 3072] doesn't match the broadcast shape [2, 8, 3072]"

[Tablaski]

Hi

Thanks for the nodes suite for this cool model

I'd like to use it with dedistilled flux models for better prompt adherence, using CFG > 1
(Or even trying with the regular flux dev model with CFG > 1)

However, setting a CFG value > 1 in the Ksampler generates this error

output with shape [1, 8, 3072] doesn't match the broadcast shape [2, 8, 3072]

I don't know how easy to fix this is but this would be great because I find the biggest issue with the model is to control the face expression, even with the controlnet, it's hard to make it really adhere to the prompt for instance "closed eyes". I've found out setting weights to 0.7 helps a lot

[EndlessSora]

@Tablaski Hi, thanks for your interest! We only tried guidance-distilled FLUX model for the moment. We use guidance in CLIPTextEncodeFlux to control the guidance-distilled CFG. We have not yet tested cfg in KSampler.