<lora:anime3d:1> will convert your anime model to a 3D style and bring new life to all of your old checkpoints. For some models it will change the style to a PVC figurine, for others it will make a pleasing artistic render with depth. This LoRA produces wildly different results depending on the model you use, so experiment!
Model used in sample image is OrangeMixs
Notes:
This LoRA can also be used as a heuristic tool. If the base model has photography blended in, it will produce more photo real results, if it has PVC models blended in those will show up, if it used 3D renders that will come out as well.
The effect should be largely agnostic, so it should respond to the underlying model style instead of applying a common one across base models.
If you combine this LoRA with your model and release a 2.5D or 3D version, I kindly ask that you link back to this page (you don’t have to though!)
Description
FAQ
Comments (2)
I recommend in negatives ((asian)), ((asian face)), strabismus, but not with a negative weight. Enhances photo prompt with positive weight.
Your LoRA works extremely good with ControlNet. I was searching for that kind of effect for years and i just needed for someone like you to make it happen. 5 stars well deserved !
A big bigggggggg thank you man ! This one is a must have LoRA.
Helpful tips:
- You may need to start with 0.3~0.5 LoRA and change it according to your seed. Starting 0.5 the 3D is obvious, so if you want less, either decrease steps or find a good number between 0.4 and 0.5 like (0.42~0.48).
-Depending what model you use be aware of the sudden seed changes especially with HiRes Fix. So you better generate seeds with LoRA first and fine-tune it afterwards.
-The more the generated image is 2D with less steps the more you need to increase the LoRA. So depending your prompts and model, finding the right middle weight to keep your prompts and adding 3D might be kind of laborious. (As the author said, go experiment :D)
-You should prevent using prompts that create realistic and 3D images if you want to use this LoRA. It works especially good with 2D, anime, manga, drawing, painting, pastel, cartoon, comics etc... prompts. You can even mix this LoRA with others, be just sure this LoRA is set at the end of the prompt chain to works better (only style-LoRA might be after this one).
-Using ControlNet with this LoRA works extremely good with txt2img using "Lineart" preprocessor and model. "Canny" works too but is a pain to prompt correctly. Many trials and errors awaits :S Interrogate Deepbooru helps but you need to fine-tune the prompts and its weights to get good results without needing to inpaint.
-Finally HiRes Fix needs SwinIR 4x at 0.25~0.4 denoise strength. Wasn't able to get good results using latent diffusion without getting weird outcomes (but i think it's model related and VAE too... didn't experiment too much models to tell you the truth).
Happy Diffusion everyone !