support me on kofi: https://ko-fi.com/n0utis
commission opened in dm or ko-fi.
you like my model? follow me!
Do you like my models? Place bids! I'll follow you!
-------------flux update--------------
I know the basemodel flux has an error, I'm trying to find a way to finetune the basemodel with standard images to clear them.
FluxGym-Kohya-Civitai
Checkpoint dedicated to anime image generation - merge approve
Original checkpoint: flux-shnell-fp8
Compatibility: Natural language & TAG (booru type)
Style: Between flat anime and illustrious
License: apache-2.0
""Created by multiple lora merges. Lora created locally and using the Civitai Lora generator.""
--Uses 16GB of RAM and 10.9GB of VRAM - 1 minute 15 seconds per frame with a basic 3060 RTX
------------parameter-------------
Vae: ae.safetensors
clip: clip_l.safetensors
T5: T5xxl_fp8_e4m3fn.safetensors
Sampling: euler / sheduler: simple
width/height: sdxl format
sampling: 15-30 steps / Distilled CFG: 3.5->7/ cfg:default
Word to improve generation quality: (perfect artistic illustration),masterpiece,ultra detailed,absurdres,high quality,amazing quality,perfect composition,incredibly absurdres,intricate details,very aesthetic,perfect hand
----------LORA training-----------
To create a Lora compatible with this checkpoint*, the basic flux model used on Civitai works, as does the original flux-shnell-fp8, fp16 & bf16 are working!
!!! Do not use this checkpoint for training, it will not work properly!!!
To train characters, use only the character's base tag* without adding anything else, such as: Hatsune Miku
*This will make it usable in both natural language and tags.
For styles or concepts, use natural language first and add the main elements manually as tags.
I don't know if this is true only for this checkpoint or all the others; other versions will be released when they are more stable and will eventually include as many characters and concepts as possible.
The next ones will be NSFW, adding as many "variants" as possible.
Several characters will be added and available natively in V2!
Please report any errors in the checkpoint, ideas, or existing concepts you would like to see added in the next version. Any help is welcome!
SD 1.5 V2 Update :
Use vae : vaeFtMse840000EmaPruned_vaeFtMse840k
Just like Easy Anime, this base model was born from multiple fusions and training. It incorporates the possibility of not requiring positive prompts for quality reinforcement; only negative prompts can potentially be useful. cfg : 7 eulera a, 40 step + adetailer
V3 illustrious update:
cfg scale : 3 euler a 30-40 step
Positive : (masterpiece, best quality),amazing quality,absurdres,
Negative : (worst quality, bad quality),bad anatomy,multiple view,Stable_Yogis_Animetoon_Negatives-neg,
enjoy the style that this basic template offers you! able to get you good quality images in a large number of samples, as long as you have low-weight loras..
integrates my loras anime enhancer v5, add detail v3
illustrous version V1-V2 :
cfg scale : 3.5-4.5 euler a, 30 step
Flux version : (tags only)
the basemodel goes in the stable-diffusion folder. it is the basemodel in the same way as sdxl or sd 1.5 it works thanks to the normal clip, you can use the vae ae.safetensor or the bf16 version which is more memory-efficient. concerning the T5, it is also the fp8 version.
if you have quality issues, use lora anime enhancer with low weight (0.1 to start)
You are solely responsible for the images generated with this model, it is prohibited by default to create illegal nsfw content or any content that is illegal or harmful to the AI community!
Using this model to train lora based on real people is not permitted.