license: creativeml-openrail-m
tags:
- stable-diffusion
- text-to-image
datasets:
- ProGamerGov/StableDiffusion-v1-5-Regularization-Images
Ukeiyo-style Diffusion
This is the fine-tuned Stable Diffusion model trained on traditional Japanese Ukeiyo-style images. Use the tokens ukeiyoddim style in your prompts for the effect. The model repo also contains a ckpt file , so that you can use the model with your own implementation of stable diffusion.
🧨 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 = "salmonhumorous/ukeiyo-style-diffusion"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")
prompt = "illustration of ukeiyoddim style landscape"
image = pipe(prompt).images[0]
image.save("./ukeiyo_landscape.png")
Training procedure and data
The training for this model was done using a RTX 3090. The training was completed in 28 minutes for a total of 2000 steps. A total of 33 instance images (Images of the style I was aiming for) and 1k Regularization images was used. Regularization images dataset used by ProGamerGov.
Training notebook used by Shivam Shrirao.
Training hyperparameters
The following hyperparameters were used during training:
- number of steps : 2000
- learning_rate: 1e-6
- train_batch_size: 1
- scheduler_type: DDIM
- number of instance images : 33
- number of regularization images : 1000
- lr_scheduler : constant
- gradient_checkpointing
Results
Below are the sample results for different training steps :
Sample images by model trained for 2000 steps :
prompt = "landscape"
prompt = "ukeiyoddim style landscape"
prompt = " illustration of ukeiyoddim style landscape"
Acknowledgement
Many thanks to nitrosocke, for inspiration and for the guide. Also thanks, to all the amazing people making stable diffusion easily accessible for everyone.