Finger Count for SD15
Ithis is not a typical LoRA you use with weight = 1. Not a hand autofixer, it is designed for only adding/decreasing finger number, it will not fix hands when they are already deformed. It will also NOT make hands more deformed (unless that's how the checkpoint model intreprets it.)
A Game Changer is here! Start a new era of no more missing fingers and extra fingers!!!💪 💪 💪
...but you can still get deformed fingers and it does not guarantie to fix all your problems
Suggested Weight: -5 ~ +5 (or higher. normally it should be 3-7 fingers per hand)
and adjust the strength according to your negative prompts as some embeddings are made to limit growth/decrease of fingers. You can use them, but LoRA weight will need to be higher to see a difference.
We all know hand is a difficult issue in SD. If you like this tool and want to see more of them, please follow me and leave a positive review. Thanks😘
This has been trained on only 14 images. if there're enough feedbacks, I will expand my dataset and update a V2.
Sample images were generated without negative prompts, blurry effect is inherent in the checkpoint model, not the effect brought by this lora. You should be able to generate better results with negative prompts.
How to use?
Positive: More fingers
Negative: Less fingers
When you have an almost-perfect image but the finger number is just wrong.
missing fingers -> Use positve weight until it gets to the right number.
extra fingers -> Use negative weight until it gets to the right number.
Extra tip: Would be better to use only OUTALL blocks.
Like what I'm up to? Feel free to treat me a Ko-fi. Cheers!
As my user name suggests, I am trying to find my NextMeal.
so I put a revised version(only relevant layers activated) on Ko-fi for supporters and it has been adjusted for more stable compositions.
It has been months with negative balance for me, because I've been struggling a lot to find a full-time job as a foreigner and I am living abroad with my partner. Your donation would really greatly help me out with my life. I am also open to commissions if you have some custom requests that I can help with, you can just leave your comments on this page or on Kofi.
Description
FAQ
Comments (27)
Anyone had this issue before? Show of hands?
I kid. Great lora!
Your sample images are actually terrible, I don't ever extra fingers or missing fingers, so what is the correct setting to always get normal gd hands?
I guess you rarely generated images with characters holding stuff. Many checkpoints avoid generating hands as it is a difficult issue. Haven't you noticed that your characters put their hands in pockets or cropped out of frame more than usual? When you say it is terrible, did you see that in my sample images, hands have issues even when weight is zero, that actually means it's not using this slider at all. This is a finger count slider, not a good hand slider. When I generate showcase images, I want to show what it is actually capable of, instead of just choosing one absolute best photo out of 100 garbages. Thanks for the comment anyway
It's easy if you think about it for a second. Generate with the lora at weight zero. If you get an image you like, but it has too many fingers, regenerate the same seed, but reduce the weight until you get the correct number. Likewise, if you don't have enough fingers regenerate with an increasing weight until you get the correct number.
You don't get missing fingers or extra fingers? How is that even possible? You must be joking.
Pretty sure he meant to say "I don't ever want extra..." not "I don't ever get extra...", that's what makes sense in context.
@xnihil0zer0254 Thank you. Your explanation is the best.
@wewewew I think you are right. but I got confused because I specifically mentioned it is proposed for fixing finger number in both the description and the cover photo. Imagine I post showcase images with normal hands, who is gonna know what it does.
@NextMeal Your "showcase" images are all over the map and typically still show deformed hands. In one example, 5 yielded what appears to be the correct number of fingers, in another it's 0, in another it's -5, and in a couple all the hands shown are completely fucked. Doesn't appear this slider does anything consistentl or any better than the mangled hands SD creates by default.
great idea! thank you
I am glad someone likes it, god knows how many hours I spent in inpainting mode. 😂😂
How did you make this? I have tried this numerous times with not luck. AMAZING WORK!
Thanks🥰 It has many steps. you can DM me on civitai discord for details.
I haven't even tried it yet but I have to express my awe. How could you have made something so amazing? This is a true game changer and I can't explain how happy I am!!!
Haha, but you still need to test it before actually knows how it works. Don’t believe everything you see on internet😂No sellers is going to say they have bad products
So, i suppose you made a slider lora with that technique to 'remove/add' concepts from a model and you set the input prompt as 'extra fingers'?
this one was trained on custom prepared dataset. 'fingers' in the substring rarely leads to good results. Like in one time I made a head size slider, I cannot type 'small head' because it has 'head' in it, and it makes things worse.
@NextMeal makes sense, I wonder if you'd made a slider lora with some long prompt about "lowres, bad quality" and such and you merged that with a ckpt with, dunno Kohya tools or Supermerger, do you think it could improve the overall quality?
@gsgsdg yes, you can do that, and such sliders have been made. you can find many of them on the site.
@NextMeal Neat! Thank you!
Oh, my God, it must be good to put it in a negative effect
just as a reminder. This is not a good hand or bad hand embedding, not a lora which should be used with weight = 1, just saying to make sure you know for sure what it does. I think a lot of people still are not familiar with using sliders. Nothing wrong with this, but right now in the gallery no one uses it correctly. Most users probably never read descriptions.
Make a joke about this a few months ago. Nice job lol
A lot of people think this is some kind of magic bullet: Put it at weight 5, it gives you 5 fingers. At weight 1, it gives you 1 finger. (At weight -1, does it rip off one of yours?)
—It does not work that way.—
There is also no way to say "At weight 1, it will generate 5 fingers, and at weight 2, it will generate 6." No. It will vary from one model another, one prompt to another, one seed to another. Just like it always has. It just nudges the image a little or a lot in the direction you want it to go. With the same model and same prompt, and seed 1, you may need to use this at weight +1. With seed 2, you may need -2. With weight 3, you may not need it at all.
You generate an image, and it has 6 fingers? Then add this lora at weight -1. Still has 6 fingers? Use weight -2. Now it has 4 fingers? Try -1.5.
Learn to use your heads and experiment, please!
Neither this nor any other lora or embedding is capable of teaching SD how to count accurately!
It works with toes, too!
I guess this can be used to fine tune the number of fingers, but does it work with drawings or just photos?
Can we get this lora for Flux?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.