nitrosocke/Arcane-Diffusion

Arcane Diffusion

This is the fine-tuned Stable Diffusion model trained on images from the TV Show Arcane. Use the tokens _arcane style_ in your prompts for the effect.

If you enjoy my work, please consider supporting me 

🧨 Diffusers

This model can be used just like any other Stable Diffusion model. For more information, please have a look at the Stable Diffusion.

You can also export the model to ONNX, MPS and/or [FLAX/JAX]().

#!pip install diffusers transformers scipy torch
from diffusers import StableDiffusionPipeline
import torch

model_id = "nitrosocke/Arcane-Diffusion"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "arcane style, a magical princess with golden hair"
image = pipe(prompt).images[0]

image.save("./magical_princess.png")

Gradio & Colab

We also support a Gradio Web UI and Colab with Diffusers to run fine-tuned Stable Diffusion models:  

Sample images from v3:

Sample images from the model:

Sample images used for training:

Version 3 (arcane-diffusion-v3): This version uses the new train-text-encoder setting and improves the quality and edibility of the model immensely. Trained on 95 images from the show in 8000 steps.

Version 2 (arcane-diffusion-v2): This uses the diffusers based dreambooth training and prior-preservation loss is way more effective. The diffusers where then converted with a script to a ckpt file in order to work with automatics repo. Training was done with 5k steps for a direct comparison to v1 and results show that it needs more steps for a more prominent result. Version 3 will be tested with 11k steps.

Version 1 (arcane-diffusion-5k): This model was trained using Unfrozen Model Textual Inversion utilizing the Training with prior-preservation loss methods. There is still a slight shift towards the style, while not using the arcane token.