- This topic has 22 replies, 6 voices, and was last updated 5 months, 4 weeks ago by
.
-
Posts
-
On this day in 1755 Giacomo Casanova is arrested in Venice for affront to religion and common decency... randy bastard. And yeh I know there's a finger issue but there was a lot of inbreeding back in those days [oligodactyly] 🙂 .
Lovely image with loads of details!
I guess your lora training is giving good fruits, or whatever the Casanova dude was doing:)
@eelgoo Proper AI training can just give a better % of getting proper hands. Feet, on the other hand, tend to give a better % of proper generation since we don't usually make complex gestures with them.
"Why is it difficult for AI to count fingers?"
There has been a lot of talk about this but basically it come down to AI finding it really hard to learn hands because they can look completely different depending on the pose, and AI is shit with complex objects.It's funny how AI can learn to pass the medical exam but not do hands LOL. I guess it's like how a child learns how to draw by developing graphic schemas like relative size, rotation, positioning, orientation, and AI hasn't quite got it yet with hands. Give it time.
@legolas18 - Maybe I'm imagining it but with hands I'm getting worse results than I did before, with he same models, prompts etc. I used to be able to use hand_yolov8n.pt with ADetailer but now it makes them worse.
The thing is Adam, and I think you've said this before, generative AI is inconsistent and very much unpredictable from one render to the next. I can get good hands by just repeatedly re-rendering but then something else has changed. What's really annoying is when you get a really nice image and only the hand/s are fucked up. Inpainting is so unreliable I don't bother with it anymore.
I havent done much AI so I speculate a lot. What little I did run is I keyed in a few things and like you said, I got different results using the same template.
I was watching southpark stuff today and this came up. Basically removing the human factor from comedy. It got me thinking about how much control people vs AI have over their image. Do they just push buttons till something cool comes up and they like it or do they have significant input on the image. 50%-90%? Or do they let the randomness of the AI tell them what they want?
@frank22 I have never gotten any good hands with Adetailer TBH, it always makes things worse.
As for hands, I am just using a good model, a DPO lora and a negatives one. But, then again, I am on the latest generation of PonyXL (I am not doing photographic stuff), which is a huge improvement on hands and feet.
I hear there is a ControlNet to fix hands too, so I'll have to look into that (not the depth technique, a new one.)
@d0gg0d
Depending on what I'm doing it will usually take 30-40 renders (sometime 100+) until I get something I think is workable. Then I have to fix the errors in postwork. For example, the left hand of the left character in the image was irredeemable so I had to get rid of it entirely, and try to fix fingernails etc. There's usually a ton of stuff I need to work on. The colours in this one was OK but I often adjust brightness, hue and saturation if it's off.
Prior to all that I'll often use an image to pose/position the characters with openpose/dw_openpose, and sometimes use the same image in ing2img as a base to work from. So yeh, it could be said I'm "stealing" other-peoples-stuff to get a result. Though I'd like to call it ingenuity 🙂 .@legolas18
Yep Adetailer is crap for hands. But if you've made a likeness lora, like the Superman one I did, you really have to use the face Adetailer for it or you won't get the likeness. Sometimes it looks too strong but I think I saw a strength adjustment for it yesterday. I'll have to play with it.
- You must be logged in to reply to this topic.
![[link]](https://images2.imgbox.com/0a/ae/AmjcZQSl_o.jpg){kind=link}