Open AnabelleKM opened 2 days ago
It's really up to the text encoder, try to use a smarter one like the VIT-l ones. if the LLM part in front of Flux doesn't figure it out, yes, that will happen.
I've found that negatives work better if you leave them out of the sensitive early steps, like:
[(tag, tag, tag):0.25]
so it's blank for the first 1/4 and doesn't accidentally introduce anything.
I've found that negatives work better if you leave them out of the sensitive early steps
This is a built-in option btw:
Settings -> Sampler parameters -> Ignore negative prompt during early sampling
Can somebody tell whether I am wrong or has experienced the same?
When adding negative prompts and then generating the image, it seems like these prompts get added to the image, instead of... not.