Can find project files(.blend) on my Patreon . Subscriptions can help me update further and create more~
!!! If you want similar results,
Please repeat most of the generate/sample settings from the raw pictures examples here,
including ControlNet(OpenPose mostly) and also others' LoRA/TI, like CharTurner | CharTurner(TI) | Baldifier slider | noline - v1.0...
different extra LoRA/TI have their special efforts, add or delete them to test what suits your targets~
Here is one of the LoRAs I trained for my new workflow, AI-powered designs, and textures(camera-based projection) tools.
They are:
- Bodyturn Helper - v3.0 | Stable Diffusion LoRA | Civitai
- Faceturn Helper - v2.0 | Stable Diffusion LoRA | Civitai
- Hairturn Helper - v1.0 | Stable Diffusion LoRA | Civitai
- Clothturn Helper - v1.0 | Stable Diffusion LoRA | Civitai
- Mechturn / Armorturn Helper - v1.0 | Stable Diffusion LoRA | Civitai
- Sword Gen - v1.0 | Stable Diffusion LoRA | Civitai
They work best with my character Library in the future,
including 3d models' mesh shape edit guide, auto texturing/shading, weight painting, and so on...~
Mostly design for asset creation to develop my own game,
which may help other independent developers like me.
Personally Recommend Checkpoint (Their versions are important too, personally use merged versions for different styles):
anime: Anything V5PrtRE Anything | animelike25DV11Pruned | AnyLoRA | Counterfeit-V3.0 | MechaDream | Mistoon_Sapphire - v2.0
realistic: Henmix_Real - v5c | AbsoluteReality - v1.8.1 | BeautifulRealistic - v1.0 | ChilloutMix - Chilloutmix-Ni-pruned-fp32-fix
Also training data here, about 30 percent of them rendered from artwork I created and posted before ,
the rest are bought and collected from the internet:
Every LoRA was trained from a relatively tiny collection including 20~ 50 pictures for this early and test version, but it still helps a lot now.
And training setting for kohya_ss for every LoRA I create, only uses different data, image size, and epoch, the rest are the same:
(recommend trying your own better data and setting,
for I am not familiar with training now,
just try to make every step/feature workable this time and keep settings as simple/default as possible,
will learn further about reasonable data, tags, and train settings later,
then train better by only using my own artwork render images to ensure it will be open to use, and follow my styles/tastes)
Here are the raw picture samples with the whole generated info that can be read
(! take care of the body turn sample, they are naked bodies and not suitable to read in public~):
Add dynamic prompt and wildcard from Civital to help get unlimited different results, using Gemini and Chat GPT to generate prompt wildcards for testing work well too~
if any questions or anything here interests you
Feel free to contact me by leaving a comment or mailing to
Also Discord Channel here~
https://discord.gg/SFJFcctfgD
Patreon Subscriptions can help to get the budget to update further, create more, and post them~