Text-to-Image Worse with This Iteration
I don't know if anyone else is experiencing this, but ever since the update to what I understand is Flux.1 from SD, the generator has been horrible about defaulting human figures to the anime-inspired, digital painting style that people generally refer to as AI slop, regardless of your style inputs.
Previously, if you specified "middle-aged women, acrlyic paining" in your prompt - with "digital painting, anime style" in your negative - that's what you'd get: a woman who looks between 40 - 50, with a corresponding body type, rendered in a fairly true to life acrylic style. Now the same prompt will get you a My Hero Academia character who looks about 12 and drawn with Clip Studio.
While () still seems to work for emphasis, it also seems we've lost negative prompting, as another user pointed out in a previous thread. If you could prompt things like, "not a teenager, not anime" etc., that wouldn't be so bad ... except that now, it seems that including either or those terms will add it to your prompt because it doesn't seem to acknowledge "not" or "no".
Is anyone else experiencing this, and know of a viable workaround?
Thanks.
for what i understood negative prompts are useless in flux so you should try to change the prompting style to work around the things you don't like,i also noticed the longer the prompt the easier it is for it to become anime so keep also that in mind
Thank you again for your tips 😊 I noticed, too, that longer prompts give more problems, and we should keep in mind that 'prompt' means the actual prompt + the artStyle description appended to it. There are some artStyles on some Perchance generators that are longer than a long prompt 😂, so it's important IMHO to use simple artStyles (like Cinematic or Professional Photo), or select "No Style" and write our own style directives in the prompt, keeping them as short as possible.
today it seems like the quality is much worse the last few weeks it has been slightly worse but today it has been total garbage. but you are talking about an update that happened over 1 year ago when perchance was way worse at generating images of anything interesting. right now it seems to be almost as bad as that version. but that was one year plus ago. over the last year the quality has been fluctuating between good and bad it seems to be different every time
Appreciate the responses, everyone.
I've tried various things to negate unwanted elements - again, use of "no" and "not" don't seem viable - but using words like "western style" seems to minimize anime influence, as an example. I don't think we'll ever get to the level of fidelity we had when negatives were active, but hopefully things will improve.