Yet Another Update 2/14/2025: retrained on new Animagine 4.0 Zero
Happy Valentine's Update
With the release of Animagine Zero for training loras, the result I'm much happier with. All example images generated on the new v4 Opt version as well.
It sounds like CagliostroLab is working to release an improved Animagine model, but hopefully this should work fine with newer releases. We will see.
Lora works well at 1.0 weight
Examples generated at 30 steps, guidance 6
Prompting remains the same as the prior release
Update 2/7/2025: new data set trained on Animagine 4.0
Retrained on the new ver of Animagine with a better data set. Honestly I still prefer the Animagine 3.1 version for the aesthetics, but this turned out ok.
Best weight: ~ 0.8 - 0.9
Steps: 25 - 30
CFG: 5 - 6
Sampler: Euler a
Scheduler: SGM uniform
Trigger: murasansei
VN outfit trigger: red dress, cleavage, bow, side slit, pelvic curtain, detached collar, black collar, pendant, detached sleeves, black sleeves, ribbon-trimmed sleeves, obi, sash, rope, side-tie panties, black thighhighs, grey legwear, ribbon-trimmed legwear, toeless legwear, sandals
Optional triggers (any one of these): dark-skinned female, dark skin, very dark skin
Illustrious and Pony version here:
UPDATE v0.2 4/24/24
No longer requires excessive prompting. Just the "murasansei" trigger word should do. See version details for more info.
Thanks!
This is my very first attempt at training a LoRA and I was kind of winging it from a guide by justTNP, so any comments, suggestions, criticisms, tips, etc. are welcome. Trained locally in one hour on a 3090 using a dataset of 101 images. I was surprised there wasn't a SDXL LoRA for her yet, so I decided to try creating one.
Trained on AnimagineXL 3.1, so should work with Animagine based checkpoints
Triggers are: murasansei, dark elf, dark skin, white hair, long hair, antenna hair, orange eyes
If you want her default look from the Visual Novel, add: red dress, side slit, large breasts, cleavage, detached sleeves, side-tie panties, black thighhighs
Weight 0.8 - 1 seems to work just fine
Thank you and feel free to post your results as well as any helpful feedback!
Description
Initial release
FAQ
Comments (4)
Hi Mr.GSone.
Thank you for always posting your cute illustrations.
I think justTNP's guide has most of the content covered.
I am not sure if this will be helpful, but I would like to mention what I keep in mind when I create a data set.
(1) Expressions of joy, anger, sadness, and happy (including ahegao)
(2) Front, back, left and right images
(3) Accurate tagging
(4) Processing of white backgrounds using Photoshop, etc. (to portray Checkpoint's own backgrounds as beautifully as possible for use)
(5) Images of the upper body, lower body, and full body (tag each one properly)
(6) Elimination of blurred images (use images that show the face as cleanly as possible)
(7) When creating a character's Lora, remove tags related to eye color, hair color, hair length, and hairstyle and consolidate them into trigger words. (Optional. You do not have to do this.)
(8)As much as possible, tags for the same element should be consolidated into one of the tags to be called. (e.g., ① "pauldrons", "shoulder armor", ② "detached sleeves", "long sleeves", ③ "gauntlets", "vambraces")
My personal favorite optimizers are AdamW8bit and Prodigy.
I think Civitai's training function is also useful there, so if you have extra Buzz, you can enjoy trying it out in your spare time.
Let's enjoy our own pace 🤗.
Thank you for the kind words and suggestions! I will keep those in mind to try and improve future attempts. And as always, thank you for your great models
Nice to meet you!
The muramasa is so well done and so cute, it's great!
So I would like to request a SDXL lora, is it possible...?
If so, I would like to request a character named Hibachi and Oharu from a work called Musibugyo.
Thank you very much!
Thank you for your message.
Unfortunately as I am just now learning, I don't think I have enough knowledge to do any requests yet. Currently I am working on improving this LoRA with a better data set, better tags, and experimenting with training settings.
I appreciate the compliment, and I hope someone is able to do those Musibugyo character LoRAs for you
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.