I generally like to keep the negative prompt at a minimum, i feel that if you put too many things in the negative prompt it will eventually remove detail from images. And as such, i haven't ever used a negative textual inversion or any complicated negative prompt bigger than 25 tags.
@nop related to hands i just have "extra hand, extra hands" and i think it helped a bit, but related to fingers i don't have anything, i just accept whatever it comes.
Some people that release models do recommend stuff like the "easynegative" lora/textual but i've used it and i don't like it's influence.
I don't know if someone has done it, but maybe the way of doing a good negative would be to make a LORA trained on BAD images and then use it as a negative, perhaps.
@ryuichi "easynegative" is good for getting different angles, but it has a significant impact on the output. Among those that improve finger drawing, I think badhandv4 is the best, but it still has a big impact on the output. Bad-prompt-v2 looks unnatural because the finger and body poses don't match. Although it has improved with model development, it's still difficult to cleanly show fingers with txt2txt.
@ryuichi@qoto.org the less is more, its been my approach to both positive and negative prompts.
@maid i remember when i started generating images with Anything V3, that model i think needed more explicit prompts to be able to get what you wanted, but nowadays these new models are more smart and you get better stuff with less tags.
@ryuichi@qoto.org The main issue with anythingv3 is that its so overfitted it just does its own thing. I have the same problem with counterfeit model. Good if you just want to roll gacha, but terrible when you are trying to do something specific. I used to merge it half and half with elysium model to tame it down.
@ryuichi Methods of inserting a large number of negative prompts related to hands and fingers, such as 'melted hand,bad hand,...' are often presented as prompt samples, but it was pointed out in a previous post that such tags are rare in the original images used for training data, so they may be meaningless. Indeed, in my experience, drawing hands does not seem to improve even when a large number of negative prompts are entered.