If you enjoyed this model, why don't you check out my other works? I put a lot of effort into correctly recreating artist styles and making the occasional character LoRA.😊
This model was trained using the collab/guide by holostrawberry [Guide] Make your own Loras, easy and free - colabs | Stable Diffusion Other | Civitai
About the artist:
Twitter: Meno Ziriath🔞🏳️🌈 (@Menoziriath) / Twitter
Patreon: Menoziriath | creating Adult art | Patreon
Meno Ziriath makes nsfw art which consequently results in an nsfw model.
Weights & Prompts
Lycoris:
This should be your goto model for most checkpoints, allowing for more flexibility.
Use a weight of 1.0, adjust to taste. For reference check the XYZ plot in the examples.
LoRA:
Use a weight of 1.0. Go a bit lower (I used 0.75) for character loras. As usual I have included XYZ plots for reference.
For backgrounds and landscapes the model is generally low on details or blurs the background entirely. It results in warmer colors, often increasing the impact of orange components. I have attached one set of landscape images with and without the model.
Please leave a review if you enjoyed this model or have problems with it. This helps me with improving future models and finetuning them in general. Give it a try and throw some images down below! 👇
Description
Trained in batches of 2
FAQ
Comments (2)
looks great
do you have any tipps for training a LoRA for a artist style? i would start doing it too (already installed and tested Kohya WebUI by creating a safetensor and it worked with Automatic1111 WebUI) but have no clue what i have to do to actually recreate the art style. Do i need more images, change specific settings or write the prompts in a different way? Is it recommended to not rescale images to 512x512 to not lose to much of the image (and with that informations for the ai, Kohya WebUI has a setting to tell the ai those to large resolution/wrong ratio)? thank you in advance ^^
At the top of the description there is a link to a really good guide. This also has the advantage of using google colab so I can repeatedly create LoRAs with my crappy GPU.
On google colab there is a dataset maker and a lora trainer. The Dataset maker can scrape images from gelbooru and sets up the necessary folders in your google drive. You can also manually download images to put them into the dataset folder. Tagging and filtering of duplicates is also done automatically but you can edit everything afterwards in your drive. I generally try to get 100+ images with little to no text. For nsfw content check rule34, they often have more images than gelbooru.
The trainer does what it says. I generally train specifically for the AOM3 checkpoint so for that you will need the link Models/AbyssOrangeMix3/AOM3A3_orangemixs.safetensors · WarriorMama777/OrangeMixs at main (huggingface.co). Sometimes just training with the anime checkpoint is good enough. Then train for 30 generations saving each with the text encoder set to 0.
The real work is finally testing through all the versions which one looks right but remains flexible! In this case generation 30 had some really fitting results but could absolutely fuck up the anatomy depending on the prompt and seed.
Just play around with the options. Everything is explained pretty well and if you have more questions don't hesitate to ask.
tl:dr
-[Guide] Make your own Loras, easy and free - colabs | Stable Diffusion Other | Civitai
- you need a lot of images for artstyles (100+) while 20 can be enough for characters
- setting are taken care of in colab
- rescaling is not necessary, images are bucketed by their aspect ratio and scaled down
- Kohya WebUI as I understand it tags images the same way WaifuDiffusion and Blip do for the colab -> you can use it as an alternative but it's not necessary
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.