Outputs differ from prompts
Hey there city96!
First, thank you for providing us those quantized models of HiDream as well as you did for us in flux.
But i couldnt get the right outputs yet. Idk why...
The prompt:a girl walks in NYC streets, with cars around her
negative: text, watermark
output for CFG 5
output for CFG 1
Anyone with suggestions?
It seems i was just hurrying to post here. But if anyone comes here with a similiar issue;
It was the custom node of QuadrupleCLIPLoader (GGUF) making this happen. I have uploaded my Comfy and found the regular quad clip loader is available now. So here is the girl and NYC:
I think the issue in this case is that scaled FP8 checkpoints and GGUF checkpoints can't be mixed in the same loader node (at least for now). But yeah, if you don't plan on using any of the GGUF text encoder files you're good to just use the base node.
Edit: I've added a check to throw an error that warns about mixing the two being unsupported, and that the default loader should be used.