You're welcome :) Enjoy!

@AshWasTaken Do you mean perfect eyes in a realistic style or an anime style? I don't really do realism styles.

@AshWasTaken I can take a look into it. AutismMix with a decent HiRes upscaler already does eyes pretty well from my experience, so it would likely have to be a in-painting LoRA which I haven't done before. But I can't imagine making an in-painting LoRA to be too difficult.

From my limited understanding, in training activation words are more or less simply a word or phrase applied to each and every single image used in training. So if the tags of every image in my training dateset start with the tag "StS-Lora", that will essentially serve as my activation tag because when you prompt "StS-Lora", the<lora:LORA:#> part of the prompt will receive that tag it was trained on and activate since it was very explicitly trained on "StS-Lora" if that makes sense. This LoRA slider did not have a singular uniform tag applied to every training image, so it's "activation" is much broader (basically anytime a person is prompted).

This is how i understand it, and I could be wrong since I have not delved deep into activation tags.

@Haiigdso3 I used AI Toolkit (https://github.com/ostris/ai-toolkit) to make this slider, and in the config for training with it you can specify the prompts you want to use to generate the images the LoRA will train off of. The script generates the images from your prompts, and then LoRA is trained off the images. From my limited understanding, since those image generation prompts are seemingly not used to train the LoRA (only the images themselves are), you don't really get the opportunity to craft an activation phrase. I could be wrong about that though, these are the first sliders I've ever made.

@gina24485404 That very well might be correct for LECO, but the AI Toolkit I used to generate this slider is a modified version of LECO that might change that. LECO is made for erasing concepts and is not really tuned for +/- sliders from my understanding. In the AI Toolkit, there is a "concept" parameter that is essentially telling the AI what it's supposed to be seeing and learning, so maybe that is akin to LECO's vector/activation phrase/word.

@denrakeiw I used AI Toolkit (https://github.com/ostris/ai-toolkit) to make this slider, which has a modified LECO tuned for +/- sliders.

Sure thing, I will look into it 👍

See my comment in the LoRAs description about CivitAI's moderation team putting limits on using this LoRA with the in-site generator.

@___marsh___ Might I ask what model you are using and an example of a prompt you're using? I did not intentionally train it with any giant/small character tags, so I would like to investigate this further.

@PurplefinNeptuna I will take a look into this, thanks. I am presume you are using one of the Autism Mixes for this?

@Gouda I don't know what they are referring to 🤷‍♂️

@green_anger Current ways to train sliders do not support training on ideas that the base model has no concept of. If you wanted to add support for something that a model has no idea of, it would be best to use a traditional LoRA/LyCORIS and not a slider.

Thanks for pointing this out, I have updated the link.

@ed_reserve532 Are you able to use any slider LoRAs with your setup? I have several on my profile you can try. This is the first I've heard of someone encountering this error with one of my LoRAs.

@upallnight I need more info for troubleshooting since I do not use Easy Diffusion myself. Are you able to use any slider LoRAs with your setup? I have several on my profile you can try.

@rockfallseveryonedies628 Can you send me a message on CivitAI with a link to this Google Collab so I can take a look? Thanks.

@tyrteds321 Thanks for the information. After looking into it, EasyDiffusion doesn't seem to be importing slider-type LoRAs correctly. I will open an issue for it on the GitHub page for EasyDiffusion.

@tyrteds321 Thanks for the reminder, I've open a GitHub issue: https://github.com/easydiffusion/easydiffusion/issues/1818

@BertLeSpurt I've never used Fooocus before. If it supports and properly import SDXL based LoRAs it should work but I wouldn't be surprised if it doesn't.

@liloko 你使用的是哪个稳定扩散(Stable Diffusion)软件？

@liloko Automatic111 WebUI 对吗？

@olternaut This isn't something I can do. The only optimization I could do is maybe try to make the LoRA smaller in case it's exceeding some user's amount of VRAM. How "optimized" LoRAs are is 99% up to the software you're using Stable Diffussion with because that's what is responsible for implementing the LoRAs into the rendering pipeline.

@olternaut It's a 40 MB LoRA which is already on the smaller size for a LoRA. At most I could shrink the LoRA to maybe 8 MB, and that would likely sacrifice quality and flexibility which is not a compromise I wish to make. I have no inclination to release an updated version of this LoRA unless something is demonstrably wrong with it that I can fix in a new version as I am a firm believe in "if it isn't broke - don't fix it".

Stock A111 WebUI is not known for being well optimized (quite the opposite actually). I personally use WebUI Forge and plan to switch to reForge in the future (https://github.com/Panchovix/stable-diffusion-webui-reForge).

@olternaut I also refuse to use ComfyUI for many of the same reasons. reForge works well for me.

Using reForge, for your reference using this LoRA to generate a 1024 x 1024 image using Eular A at 40 Steps, 2x Hires Fix with 4x_fatal_Anime_500000_G, and CFG Scale 5 takes about 37 seconds for me. If I remove the LoRA is take about 36 to 37 seconds, so the difference in negligible.

@olternaut What commandline arguments are you launching reForge with? You can find them in the webui-user.bat folder. Here is what I use:
--listen --enable-insecure-extension-access --xformers --pin-shared-memory --cuda-malloc --cuda-stream

I'm curious as to what your render times are with these with your 3090.

@olternaut I took a look at that video and he doesn't do any of the commandline optimizations. Since you have an RTX card, you're missing out on performance by not using --xformers in your commandline args in webui-user.bat . I would highly suggest you add "--xformers --pin-shared-memory --cuda-malloc --cuda-stream" to your commandline args and try again.

@Halfaron When you say change distance, do you mean the distance of the "camera" to the subject? AKA, how close or far the subject is in the picture?

@Sterned I have not used regional dividers before, but this sounds like something in the SD tool you're using isn't properly separating the prompts. To the best of my knowledge, there isn't any reason why this LoRA would be overriding regional prompting if it was implemented correctly.

@megachina31296 This LoRA should have little to no impact on style, so you should be able to use style LoRAs along with the age slider without any issues.

@Masked_anubis I raised this issue on the EasyDiffusion Github page months ago - https://github.com/easydiffusion/easydiffusion/issues/1818 . If they haven't fixed it by now, they probably never will. I would suggest migrating to WebUI Forge anyways because it's more performant and (imo) a better interface than EasyDiffusion. I also don't think EasyDiffusion is getting updates anymore as the last update on the Github repo is 2+ months old.

@IlyaEfimov I would if I could. As stated in this LoRA's description, the CivitAI Mod Team disabled the use of this LoRA with the in-site generator because of CivitAI's safety policies. It's not something I can change.

@Janet Last I checked, the tools to do so have not been made yet.

@Janet I'll take a look into it, thanks for letting me know.

@martinjas393 How to do what?

@MrShouttaKun Have you tried adding the prompt small eyes or similar?

@MrShouttaKun This slider LoRA was trained exclusively on an anime/cartoon dataset, so honestly I'm very surprised it works as well as it does on different style.

@rsngo What are you using to run SD? A111 WebUi, Forge WebUi, ComfUI, SwarmUI, or something else?

@rsngo fooocus is not something I've used before. I'd recommend trying out SD Webui Forge - https://github.com/lllyasviel/stable-diffusion-webui-forge

@BigSad11 Some of the photos reference "v4" because they were made with the fourth version of the slider I trained locally. However, versions 1, 2, 3 were test iterations and never released. For the public release I renamed v4 to v1 because it is the first public release of the age slider. My "age_slider_v4" is the same as the public "StS_age_slider_v1". In the example images, the LoRA names that start with "StS_" are this renamed LoRA. "age_slider_v4" is from before I renamed the LoRA file.

For my new LoRAs, I tend to use "Iter" (for "Iteration") instead of "v" now to not cause this confusion.

@Marissa This is part of PonyXL. This article explains it: https://civitai.com/articles/4248/what-is-score9-and-how-to-use-it-in-pony-diffusion

@ninjahatoori8 I would if I could. As stated in this LoRA's description, the CivitAI Mod Team disabled the use of this LoRA with the in-site generator because of CivitAI's safety policies. It's not something I can change.

@ninjahatoori8 It's possible CivitAI might allow a uni-polar "Age Up" slider, but I find it unlikely. CivitAI very much likes to err on the side of caution. I may make one regardless to test out some new training methods - we'll see.

@navy123 Currently working on it - training the LoRA as I type this.

@avicdar410 That is possible to do, yes, but why do you need it +/- 2 instead of +/- 5? Doing +/- 2 while keeping the same support age range would be quite difficult to balance and keep the LoRA from jumping too much between input values.

@dita That is a fair point except for the fact that this LoRA can not be used on CivitAI's on-site generator as it has been blacklisted by the CivitAI moderation team.

@vicentlennie956 I don't understand what you mean by this given the context of this conversation.

@dita The preferred weight values is simply stored in a JSON file in the same directory as the LoRA and with the same name. It has a "preferred weight" attribute that you can edit to whatever number you please and it will work. Example: https://i.imgur.com/frw9k5H.png

That being said I am planning on replacing my old LoRAs with Scaled ones for online services that supports +/- range. Not terribly hard to do.

@rogue8888992 So basically a hip width slider or a body proportion slider decoupled from age?

@rogue8888992 That's an interesting idea. I will add it to my list of LoRA ideas to look into.

@rogue8888992 Currently have Breast Size Slider for IllustriousXL that I just need to create cover art for before publishing. Besides that, currently working on semi-custom tools for better slider training so I can create sliders with lower inherent noise offsets and better disentanglement. If successful, I will like update my IllustriousXL Age Slider first with this new training method.

@ronikush While I can't give away all my secrets, feel free to contact me via Discord: shed_the_skin

I eventually want to update my training toolkit to support Chroma, and having someone with access to a 96 GB GPU to do a test run or two would be very helpful if you want to help. Otherwise, no worries.

@olternaut Downloads are enabled as long as you're signed in. If you're having trouble you can also download it from Sea Art - https://www.seaart.ai/models/detail/e7910d18e321122d626db1ba5def109d

@olternaut I already did. Check my profile.

@BatesianMimicry Looking at CivitAI's ToS and Safety policy, theoretically it should be possible to do this. However, the messaging I have received from CivitAI Moderators indicates that they do not want any age adjusting LoRAs to be used with CivitAI's generator service. This leaves me not very interested in putting in all the effort required to make an 'Age Up Slider' only for it to likely be removed by CivitAI.

@Gr1ef If you want finer control of a known character's age, an age up slider may be helpful - especially if you're also using a character LoRA. I can think of several situations where aging up a character may hard to do without a LoRA.

@79513685587k937 PonyXL is based off of Stable Difussion XL, which does not support per-subject prompting like FLUX can. To be able to control individual characters in an image with PonyXL you need something like the Regional Prompter extension or you need to use inpainting.

This shortcoming is not something I can fix.

@astarotte I am not sure what you mean. As far as I can tell, it's labelled for PonyXL. If you're using some sort of automatic metadata reader, it may have gotten it wrong.

@SaburoDio With over 5,000 positive reviews, I am rather confident it does work. Can you share your workflow so I can help you get it workig?