Use a different captioner (for Flux)

Although the captioner for Flux natural-language is decent, it tends to say things like “probably a x or a y”, and it often will include something to the effect of “a high resolution photograph of” which doesn’t seem helpful.
Although in Beta, I’d suggest trying Joy Caption Alpha #2 instead. Although it’s used for NSFW, it allows model instructions that would keep it from that if needed. It also addresses other issues by having AI flags that keep it from mentioning resolution, ambiguous terminology, and other such things.
I would also suggest allowing us to use empty captions, something you guys even recommend in a Flux lora training article, though when I’ve tried empty captions the model failed and said it needed them (so I’m assuming your online trainer currently doesn’t support it).