Weights can swing very far on this one -8.0 to +8.0. It can do extremely large crowds the higher you go and I wanted to be able to keep granular control.
Positive = More people
Negative = Less people
Simple LoRA to help with adjusting the number of people in a picture. You can swing it both ways pretty far out from -8 to +8 without much distortion.
Description
FAQ
Comments (43)
Nice work !!!
Incredible idea !!!! You are amazing bro !!!! 5 STARS FOR YOU !
thank you!
Agreed - this is genius!
fingers on a hand slider when? jkjk
don't think I won't :)
@Ostris That would be awesome. SO sick of getting six finger monsters and three finger aliens...
If this works you've basically solved AI.
So I ran this through my slider training algorithm just to see what would happen. The samples would sometimes add or subtract a finger, but then it would go back to 5, 5 perfect fingers. Not warped or distorted. Just normal hands. And it is doing it on both sides of the slide. The two sides are completly divergent and are making drastically different images now, but those images are pretty consistently generating good hands.
I think there is something to this. I am going to keep working on it
@Ostris I swear if you figure this out you deserve some sort of reward. Like the biggest issue with AI is the deformed hands.
@crawlingrat Nothing less than a nobel peace prize!
This is amazing!
I left you an email. When you find time please have a look at it. I will highly appreciate your response
thanks, you have create many useful tools for tuning the image easily.
What's your training method?
I tried to make a "slider" LoRA, but I did not get very good results. I would love to pick your brain or see your process @Ostris
I added links to my GitHub repo in the meta on the newer models, but it is not quite ready for public use yet as there is no documentation or config files and the read me is wrong. I am also making massive breaking changes directly to the main branch every day. I’ll try to get that updated today and maybe do a small tutorial if I have time.
Ah nice! Thanks for the hard work! I'll hold off until you feel like it's more presentable :)
Will you make a blog post or whatever here at civit when it's ready? As a way to get a notification for it ^^
@mnemic will do
Your LORAs are turning out great! Do you think it is possible / achievable to make a LORA that removes children / minors?
there are several already in exsistence...one i remember one called asyoungv.2 + asyoung negv.2 or similar try a search
I have an age slider that does this at positive weights, I am just hesitant to release it as the negative weights produces the opposite effect but not contextually aware the way prompting and embeddings are. Meaning, if your character is doing adult things, it will still be doing those adult things, just as a child. Not sure if I want to put that out there.
@Ostris i have aged sliders in use, truth is what´s good for one side, is maybe bad for the other side. In any case you could as a moderator...
I'm using your Weight, Muscle and Skin Tone sliders and they are all excellent, combining the weight - and muscle + is insane.....Great work.
Guessing the Hair and People are also must haves...
This is very useful, though there is an issue that people in the background often are facing away from the camera, and they frequently do not match the description I gave in the prompt. That's fine if I want varied people, but if I want them to look similar it is frustrating.
multiple people is a problem in general, as you probably figured it out by now. You either inpaint with a different prompt and seed or you might use regional prompt or something similar.
actually the finest and best trained lora of its type, superior in all aspects. would rate 6 stars if i could
Sorry if this is a weird question, but how can you put the slider more or less than +-2? I can reach only those two numbers on the Lora slider bar
Type it in.
<lora:people_count_slider_v1:6>
@dita thank you!
If your software limits you, like Draw Things, you can add the lora multiple times.
thank you for this LORA, now is getting much more easy to generate multiple people image
This Lora has an interesting side-effect. When you don't use the added 'more people' prompt, it can often make the scene look more complex. Words will even influence clothing and hair more for instance. That alone makes this my favourite Lora.
I'm working with the azovya's RPG artist tools model and for some reason, this LoRA seems to want to try to turn my character female so often. I narrowed down every single part of the prompt, and this was the cause. I do not know why.
It is an issue with the bias in the models, especially the ones the community has tuned. Most community trained models, if you prompt "person", and generate 10 images. You are going to get a person who is female 9 (if not 10) of those generations. This LoRA simply targets the models understanding of the concept of people and amplifies (or reduces) that concept. Any bias is also amplified. I could attempt to counteract that bias, but that would require adding a male bias, which would be hard to tune, and would only generate males on a model that has no bias.
Another option is I could make 2 of these, one that only amplifies the number of males and one for only females, and you could tune each to add more of the gender you want, or do them together to balance it. Would that be something you would be interested in?
@Ostris Given how Stable Diffusion's crowds tend to be the same person over and over with tiny variations, this by itself could do a lot to create more realistic (or at least more interesting) crowds.
@Ostris Well, the idea of it being tuned to a specific gender that way does sound good, but I don't know if I would use it enough to ask it of you, lol.
hi dude, could you make a XL version for this lora, plz！thank a lot~
What is an XL version? What does that mean?
@ModelHub for SDXL
@BIG_A what is SDXL? Sorry, I'm new to this
@ModelHub I'm ten months late, but let's do this. There are currently 3 primary model bases in use: Stable Diffusion 1.5 (SD1.5), Stable Diffusion XL (SDXL), and Flux. Also, a finetune (more on that I'm a moment) of SDXL, called PonyXL (or just Pony), is often considered as it's own base model.
Finetunes are where someone took one of the base models and gave it additional training (by supplying it with many additional photos, usually of specific things, with descriptive labels) to teach the model certain concepts that it may not have a great understanding of. For example, using SDXL, someone might notice it's particularly bad at creating amateur photography. They then take a few thousand (or many more) amateur photographs, label them, and train the base SDXL model with those images. They do this on powerful computers, and for a long length of time. Once done, they test the model to see if it can produce better amateur photographs (and if they've done it correctly, it will).
This website is basically a connection of different finetunes. There are, of course, other things, too, such as LoRAs and embeddings, which are both ways to expose ideas/concepts to a model without having to go through the hassle of finetune training. You could create a LoRA with only a few dozen images, in a short duration of time.
When the numbers are high, everyone's mouths tend to hang open.
Is it possible to get a people slider for pony models? Thanks.
Please can you upload this to Tensor art 🙏 such a great model
Details
Files
people_count_slider_v1.safetensors
Mirrors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
人数滑块_v1.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
S_people_count_-05-5.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
people_count_slider_v1.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.