tldr:
switch your prompting style, stop using negative/positive TI/loras, dont use schizo negs
Switch Your Prompting Style:
The Text Encoder doesn't like commas, at least how we use them. The Text Encoder was trained on more English text rather than tags. If you use commas a lot, then the Text Encoder gets confused, since it sees one token a lot more for no discernible reason.
To fix this, just get rid of the commas! The Text Encoder will confuse the tags since there's no difference between tag name and different tags, but we can fix it by replacing the spaces in tag names for dashes. If you replace the spaces for dashes and commas for spaces, it becomes much better at understanding the prompt.
(it also saves around 3 tokens per 50 tokens, so that's neat)
Stop using negative/positive TIs or LoRAs
I have yet to find a good negative TI/lora that doesn't bleed into my images. Most/All negative TIs will have some sort of bleed-in. In other words, depending on the data used for training, a certain style will always have to enter in. (unless it is trained exactly on the data used for the model but in that case it would do absolutely nothing) [see addendum for images]
Don't use schizo negatives
Every time I see people complain about their images being Not Good, I see that they have absurdly long negatives. Absurdly long negatives do not help images any more than simple negatives. Stop using lowres, ((bad anatomy)), ((bad hands)), text, missing finger, extra digits, fewer digits, blurry, ((mutated hands and fingers)), (poorly drawn face), ((mutation)), ((deformed face)), (ugly), ((bad proportions)), ((extra limbs)), extra face, (double head), (extra head), ((extra feet)), monster, logo, cropped, worst quality, jpeg, humpbacked, long body, long neck, ((jpeg artifacts)), deleted, old, oldest, ((censored)), ((bad aesthetic)), (mosaic censoring, bar censor, blur censor) and use (low-quality worst-quality:1.4) (bad-anatomy) (inaccurate-limb:1.2) bad-composition inaccurate-eyes extra-digit fewer-digits (extra-arms:1.2) .[see addendum for images]
Addendum
All images generated with same seed, sampler, positive prompt, step count, etc...
Negative TI's
No negative TI:
With negative TI:
Schizo Negatives
With good/short negative:
With schizo negative:
