Update: After some testing, it seems that the bug with DF11 Klein outputs in ComfyUI is limited to situations where CFG is enabled, meaning that the models without "base" in the name are not affected since they are not meant to be used with CFG, including this version that uses KV cache. I will continue to look into this issue, sorry for the inconvenience.
Note: It appears to me that the DF11 FLUX.2 [klein] models are bugged, often producing outputs that are slightly different from BF16 outputs. I have confirmed that the DF11 weights are indeed 100% identical to BF16 weights when decompressed, so this should be some bug with inferencing that I am struggling to diagnose. In the meantime, I recommend using the this model with the default KSampler node, instead of the CustomSamplerAdvanced node. Also if you have the VRAM to run this model in BF16, I recommend to use the newly added DFloat11Decompressor node to load this model instead. This node fully decompresses the DF11 weights into BF16 weights, and outputs a model that is 100% identical to the Load Diffusion Model node. I apologize for the inconvenience, as I investigate this issue.
