Has anyone experimented with secret annotations in prompts?
In older SD models, we could include things like \[some message::0\] in prompts — text that would be ignored by the model but still included in the raw input. This allowed us to embed metadata or trigger filters without affecting the image generation.
For example, we used \[ContentLabel::0\] to signal age restrictions. Our filter system treats this as a PG-13 banned prompt keyword, so users could self-label their content by adding it.
However, it seems that in newer models, this trick no longer works. Possibly because prompt weighting syntax like :: is no longer interpreted the same way — or ignored entirely.
Did anyone find a working alternative? It’s important for us since we rely on this system to label and filter user submissions.
Since it doesn’t use the previous prompting, then it wouldn’t work anymore. Anything in the prompt is taken into account in generating the image. I also don’t think there is any Flux prompting shenanigans (or hidden/secret annotations) included in the new model that can hide hidden messages on the prompt (as far as I’m aware).
Is the dev with us right now 🫨
I think some label that is displayed in the img data (label:::Label) would do the trick (but it can only be made in the text 2 image plugin i think), . For now you can use the negatives to label and filter with bannedNegatives, they affect less or nothing.
I tried ContentLabel, SFW, PG13, MetaNonVisual in positive or negative prompt with seed control - the impact of ContentLabel in negative seems most subtle
