Creating models is a labor of love, but it takes a significant amount of time and compute power to get them just right. If you’re enjoying my models, consider fueling my next project with a coffee on Ko-fi ☕. Thank you for keeping this project going!
AnimeBoysXL V3
Features
✔️ Good for inference: AnimeBoysXL v3.0 is a flexible model which is good at generating images of anime boys and males-only content in a wide range of styles.
✔️ Good for training: AnimeBoysXL v3.0 is suitable for further training, thanks to its neutral style and ability to recognize a great deal of concepts. Feel free to train your own anime boy model/LoRA from AnimeBoysXL.
Inference Guide
Prompt: Use tag-based prompts to describe your subject.
Tag ordering matters. It is highly recommended to structure your prompt with the following templates:
1boy, male focus, character name, series name, anything else you'd like to describe, best quality, amazing quality, best aesthetic, absurdres2boys, male focus, multiple boys, character name(s), series name, anything else you'd like to describe, best quality, amazing quality, best aesthetic, absurdresFor more detailed documentation, you can visit my article on Ko-fi (available to supporters only).
Negative prompt: Choose from one of the following two presets.
Heavy (recommended):
lowres, bad, text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifactsLight:
lowres, jpeg artifacts, worst quality, watermark, blurry, bad aesthetic
VAE: Make sure you're using SDXL VAE.
Sampling method, sampling steps and CFG scale: I find (Euler a, 28, 8.5) good. You are encouraged to experiment with other settings.
Width and height: 832*1216 for portrait, 1024*1024 for square, and 1216*832 for landscape.
Training Details
AnimeBoysXL v3.0 is trained from Pony Diffusion V6 XL, on ~516k images.
The following tags are attached to the training data to make it easier to steer toward either more aesthetic or more flexible results.
Quality tags
best quality: score >= 150amazing quality: score in the range of [75, 150)great quality: score in the range of [25, 75)normal quality: score in the range of [0, 25)bad quality: score in the range of (-5, 0)worst quality: score <= -5
Aesthetic tags
The aesthetic tags of AnimeBoysXL v3.0 reflect my aesthetic preference.
best aestheticamazing aestheticgreat aestheticnormal aestheticbad aesthetic
Rating tags
sfw: generalslightly nsfw: sensitivefairly nsfw: questionablevery nsfw: explicit
Year tags
year YYYY where YYYY is in the range of [2005, 2023].
Training configurations
Hardware: 4 * Nvidia A100 80GB GPUs
Optimizer: AdaFactor
Gradient accumulation steps: 8
Batch size: 4 * 8 * 4 = 128
Learning rates:
8e-6 for U-Net
5.2e-6 for text encoder 1 (CLIP ViT-L)
4.8e-6 for text encoder 2 (OpenCLIP ViT-bigG)
Learning rate schedule: constant with 250 warmup steps
Mixed precision training type: FP16
Epochs: 40
Changes from v2.0
Change the base model from Stable Diffusion XL Base 1.0 to Pony Diffusion V6 XL.
Revamp the dataset's aesthetic tags based on the developer's preference.
Update the criterion of quality tags.
Use FP16 mixed-precision training.
Train for more epochs.
Special thanks
chefFromSpace for his assistance with the showcase images.
License
Since AnimeBoysXL v3.0 is a derivative model of Pony Diffusion V6 XL by PurpleSmartAI, it has a different license from the previous versions. Please read their license before using the model.
Description
Change the base model from Stable Diffusion XL Base 1.0 to Pony Diffusion V6 XL.
Revamp the dataset's aesthetic tags based on the developer's preference.
Update the criterion of quality tags.
Use FP16 mixed-precision training.
Train for more epochs.
FAQ
Comments (94)
Yay it finally released~ Gonna try it now! Thanks as always XD
Oh look what I've found! A question before trying it, do we need the score tags and/or pony style loras for this?
@hello_world You don’t need the score tags or any style loras for this. But you can still use them if you prefer those styles!
Great model, finally some checkpoint based on pony centered on males. Just some observation: shouldn't the base model be listed as pony here on civit ai?
Thank you! I've updated the base model category.
As always, I'm thrilled to see you release a new model, Koolchh! Queria told me this was here, so I ran over as fast as I could to grab it. 😁 It's downloading now. I'll compare it with base PDXL and AutismMix. I'm expecting I'll like this one more than either, knowing your track record for excellence. Thanks for making and sharing. 🧡🧡🧡🧡🧡
And, finally, there's a great 2D model that is trained very similarly to my tastes and preferences. I'm ecstatic!! Of course, I have used v2 also, but this looks to be 5x (or more) better.
@dita I'm so glad you're excited about this! It brings me so much delight knowing there are people like you find my model aligns with their preferences so well. Thanks again for your kind words and for being such an awesome supporter. Enjoy exploring the new model! 🧡
@Koolchh It's always good to find people with similar tastes and interests. You love and I love it.
I played with this model a bit today, and it seems to behave quite differently from both base Pony and Autism. With those two, if I use the same prompt, seed, and settings, the images are fairly similar. Not so much with this one. It's very different from those two. Not in a bad way, though. The Styles LoRAs that prgfrg23 posts also behave pretty differently with this mix, so I'll probably treat it in my head as being a distant cousin of Pony instead of a child.
Of course, only one day doesn't really give me ample time to judge it. After one day of playing with Pony, I was ready to trash it because I couldn't get anything but crap out of it. But now, Pony and its derivatives are all I've used for the last 2 months. Not joking. It responds to prompts the way that models should have been responding all along. So I will play a lot more with this model. A male-tuned Pony is like a dream come true, you know? How could I not be excited about it?! 😁
@Koolchh So at first I was a bit underwhelmed with this model, and I was a bit stingy with my praise. The results I was getting were pretty bad tbh. And then I tried prompting it like you suggest instead of like I do with Pony/Autism. And that made all the difference! If I'd only read and followed your directions in the first place! 🤪
Since I started using it the correct way, it's clear that it blows away all others. I believe I'll be using this model for a long, long time. Other than very specific cases where I need something different, like a photorealistic model, it's probably the only model I'll be using for the next several months at least, unless something comes along that really blows ABXLv3 out of the water. That's a pretty tall order to fill, though. Everything about this model superb! ❤️🧡💛💚🩵💙💜🤎🖤
@dita I see. This model need more careful prompting than my previous ones! I just post an advanced prompt guide for AnimeBoysXL V3 for my Ko-fi supporters. It has more detailed information on how to generate characters. I think you'd benefit from it a lot.
Awesome!! thank you very much!! so does Lora, who has trained before, need to retrain?
They work, but the quality tags can overpower the loras so you may need to tone down on them when using loras
You mean the loras trained on AnimeBoysXL v2? I think they need to retrain since v2 and v3 differ a lot. If you mean loras trained on Pony v6, then I think they should also work on this.
Oh true, I forgot there was a previous version of this, in which case I have no idea but Koolchh is right, my response was in regards to pony loras 😅
@Koolchh Ok thanks
oh my gosh! my eyes are so blessed!
thank you for making this! i cant wait to test it!
Love this one! Btw I have a question, in an image of 2 characters, is it possible to make them look how we want? For example in an image where one person is sitting and the other one is standing, and if I want them to have different hair colours and characteristics or use different character loras on them, is it possible to do that?
try installing this extension in stable diffusion webui-forge: https://github.com/Haoming02/sd-forge-couple
@ivorysky Thank you
Great model, so far my favorite.
Thanks for sharing.
Great model, but I dont know how to use XL. Any help?
@tpalnm248 For SDXL models, you should not use a 512x512 resolution. Please use one of the following resolutions: 832x1216, 1024x1024, 1216x832.
After a simple test, I found that the character's eyes seem to be prone to errors and cannot be as perfect as the preview picture here. How can I debug them?😢😢
I use adetailer. It's an amazing extension which can automatically inpaint the face for you.
@Koolchh Got it, sir! let me study🫡
@Koolchh I tried to enable the lora plug-in, but it doesn't seem to work. How do I need to adjust it to make it effective?🤧
@Koolchh I tried the lora model trained based on pony and it didn't seem to produce a change.🤥
@Sunken Did you include the lora in your prompt? Go to the lora tab and click on the lora, it should show up as <lora:lora_name:1> in the prompt.
@Koolchh 😧 ! ? The model I used before will take effect if it is activated directly. No prompt words are needed. Do I need to add prompts to this model? Let me try it.
@Sunken For automatic1111 web UI, you don't need any plug-in for using loras; it supports loras natively. As long as you put the loras in the correct folder (models/Lora by default), and follow the steps I just described, you should be able to use them.
@Koolchh Finally debugged successfully, oh my god. I haven't used it for a long time, and I haven't tried many new things. For example, the old lora training group can no longer be applied to the pony model, and there are many usage specifications. Pony is really new.😇😇 So tired, so tired, so tired🤕
@Koolchh It seems that the full body effect is difficult to handle, and the face is always easily distorted. Sad😂
Is there a list of what characters are built into and ready to be generated with the checkpoint I can look through anywhere?
I think the list characters are the same as the original pony one, you can also use lora.
@zaveid I have noticed it can do more characters than base Pony through some testing, Squall from FF8 for example. I am sure someone, somewhere, helped in adding them. It is called "AnimeBoys" after all.
@zaveid It knows a lot more than pony, pony knows very little of male characters, @gtgiygus your best bet is to browse danbooru and check how many images a character has, the more it has the more likely it is that it can be prompted
@Queria What I have been doing is "guess and checking" by typing a name, series, and hair/eye color. But browsing through Danbooru might give me a character I had not known of to be thirsty over I suppose.
this is unironically the best anime man checkpoint model
God I love men...
the best model for anime boys, very little mutation/ugly with the correct tags, beautiful art style, good quality images, can't find anything wrong with it, a variety of popular characters are included
I've been testing it extensively these past few weeks and it feels like it started to "overtrain" somewhat, which happens when details are WAY too 'sharp' plus when brightness and darkness are too exagerated!
Could you please merge this model with 0.10% of sdxl_0 only on the UNET, and then do a quick finetune on the same dataset or another one with similar distribution so that it re-heals the information while still having more range?
sdxl_0 is a completely random-weights model that can be found on huggingface, it is useful to merge tiny percentages of it on the unet when you want to kind of "rebalance the model" due to it slightly destroying the information contained before, according to how much you merge with it.
You, sir/ma'am, you created the one model to rule them all. I kneel before your magnificence.
Words really can't convey how great this Model is. Not only can generate many male characters even without LoRas, the generated character even has their "Personalities" intact. Unlike other anime based model I've seen in this website!
For example, look at my Gudao x Achilles, Izuku x Bakugo, and Hiiro x Rinne pictures I generated.
You can see that Gudao & Hiiro is "Nice boy", while Achilles & Rinne has kinda cocky personality, just like how they are!
I can't imagine how can we get any better from this! Applause to the creator~!
I love this even much more than pony~，thank u bro~ do u have plan to make V4 version? the color of V3 is a little bit strange...
Great to hear you're enjoying it! I'm still putting in the work to take this model to the next level. Now that I've tackled the low-hanging fruit, it's going to take some more time and effort to get the future version just right.
@Koolchh It's fantastic to hear about your progress and dedication to refining the model. Looking forward to it!
Hello, I want to ask. Can this AnimeBoysXL model make characters do specific activities such as running, swimming, walking, doing push up, etc.?
its super great but there is a huge issue with faces, theyre almost always distorted, pony doesnt have this issue for some reason, but at the same time have fun getting 2 different characters in one photo there (with loras and stuff)
The faces are fine here, maybe it's your configurations. Did you follow the recommendations proposed on the description?
Hi, I noticed you trained the model with different art styles of some artists. Which ones did you use. These are the art styles I found so far (if you guys find more, please add):
- Blake Alexander
- bludwing
- dizdoodz
- hiroki-bzdsk
- kakenari
- kukumomo
- mentaiko
- robokeh
- schizoid
- suyohara
- vinceteaa
- zifu
(list so far...)
Thanks, I also found laovaan, lvlv, and mazjojo
Can we choose the style if we add the artist name in the prompt?
@Kaze111 Yes, that's the point!! :-) I'd like to know if you people discovered other artists beyond this list I wrote.
this is helpful! <3
Please turn on "online generation" in model settings, I would try it onsite, cheers
All my models should have been available for use through Civitai's generation service, as I have permitted through the settings. However, it appears that Civitai has discretion over which models are ultimately selected for inclusion in their service, and unfortunately, my models were not among those chosen.
@Koolchh that's super sad, I've been hoping to incorporate this in online generations
How does the year tag affects the prompts?
It's a way to control the style of the output. Using year tags or not is totally up to you, but `year 2023` would generally enhance the aesthetics.
can we please have this model online generation?
This is the most superior model I've experienced thus far. I'm looking forward to version 4 being made in the future .. ❤️
Very nice model but it can get better at background
One could try using loras for background. Or try using other checkpoints for the initial spep and finish with this model on hires.fix or refiner or both, or viceversa
This model is incredible. I'm looking forward for the version 4. I have some improvement ideas and constructive criticisms for your next movement if you wanna listen <3
The hands: Sometimes there are 5 fingers or 3 fingers. No negative prompt or Hires .fix or inpainting tool can fix that 💔 Just photoshop.
The hand nails: Sometimes the nails are half-drawn or poorly drawn. I simply remove the nails after generating the images and everything is ok. That's the advantage of creating anime images, sometimes fingers don't have nails and that's ok. It would be interesting to simply remove the nails or make the model learn to draw them better.
Bara characters: This model is amazing when drawing soft and/or young (twink/twunk) characters, however when drawing bara ones it loose a bit quality on the output. For example, Dragon Ball Z characters are not good for this model (goten, goku or trunks for example), this problem occurs even if I use their respective available LoRAs.
These are just minor criticisms and improvement ideas for the next version. It would be awesome if you work on that on the next one. I'm willing to help if you want. Your checkpoint is hands down the best I've seen here for a while, I'm loving that. Cheers ✨
I have the problem that in almost every image there's text on the bottom right or left and I can't get rid off it even with multiple negative prompts. any advice?
try putting "watermark" in negative
Please use this.
I erase all credits for the images I generate with this.
Look up how to use it yourself.
https://github.com/aka7774/sd_lama_cleaner
Hi, you can download proper adetailer model to get rid of those almost automatically. E.g. https://civitai.com/models/699636/patreon-watermark-detector-yolov8
@NikPureSilver Thank you so much!
@PlumBeerThank you!
大佬什么时候更新大模型啊，AnimeBoysXL V3人物太爱了，就是颜色总是不对劲，颜色怪怪的，发虚，发灰，期待新的模型
负面提示词放monochrome
Is there a way to reduce the number of abs? This model, specially when refined, produces too many pairs of abs divisions.
Great model so far, testing v30 rn.
it doesn't always help but I get a lot of better results when I use the prompts "twink", "slim athlete" and maybe "six-pack abs"
As an avid user of AnimeboysXL I want to point out that the sampler DPM++ 2M Karras works very well and, in several cases, gives HD results with very little effort. Perfect for more realism
Thanks
I'm having problems to use your model because it's keep saying "Remix
Some resources used to generate this image are unavailable" can you fix it?
Aside from what everyone else has already said, this model also has a really nice composition. You can really easily create very dynamic scenes with great lighting, angles and gesture. The only thing that it lacks is detail, both in the style, the characters and the background. I've been dealing with this by using loras or refiner with a model that is good at details and style adherence like Midkemia.
One of my favorite models by far.
Don`t you planning to create Illustrious version of this model? Because it`s absolutely amazing model, the best of all I use <3
NoobAI version please?
if you make an Illustrious version of this, it will be an awesome Christmas present for everyone ❤️❤️
Do it! How you are going to do this? Which images are you choosing?
This is still the best checkpoint arround here.
The world needs an ILXL version.
agree
This checkpoint had illustrious quality in pony
Hoping for it to come back
An IL version of this GOD checkpoint would be so cool! I'll wait however long it takes!
I'm waiting for an IL version of this model to start working with Illustrious as a whole!
Please do it 🙏 🥺
Supporting this!!!
it's been a year daddy. i really really miss you 💔
Imagine a version of this model, based in NoobAI.
Built on the incredible NoobAI base, AnimeBoysNabla ∇ is finally here! I’m so excited to share this with you after a lot of intensive training. If you enjoyed AnimeBoysXL V3, I really hope you’ll love this one too! ❤️
oh my god. i cant believe i missed this announcement!!! you just made my night! <3
thank u very much for all the work that you do!
i'll donate soon!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.