Trinart Stable Diffusion V2

🚀 Stable Diffusion TrinArt/Trin-sama AI finetune v2

This is an improved version of the original Trin-sama Twitter bot model, aiming to retain the original SD's aesthetics while nudging towards anime/manga style.

🚀 Quick Start

This model is NOT the 19.2M images Characters Model on TrinArt. It's an enhanced version that tries to keep the original SD aesthetics and lean towards anime/manga styles.

Other TrinArt models can be found at:

• naclbit/trinart_derrida_characters_v2_stable_diffusion
• naclbit/trinart_characters_19.2m_stable_diffusion_v1

✨ Features

Diffusers

The model has been ported to diffusers by ayan4m1. You can run it from different branches:

• revision="diffusers-60k" for the checkpoint trained on 60,000 steps.
• revision="diffusers-95k" for the checkpoint trained on 95,000 steps.
• revision="diffusers-115k" for the checkpoint trained on 115,000 steps.

For more details, check the "Three flavors" section.

Gradio

We support a Gradio web ui with diffusers to run inside a colab notebook:

💻 Usage Examples

Basic Usage

Example Text2Image

# !pip install diffusers==0.3.0
from diffusers import StableDiffusionPipeline

# using the 60,000 steps checkpoint
pipe = StableDiffusionPipeline.from_pretrained("naclbit/trinart_stable_diffusion_v2", revision="diffusers-60k")
pipe.to("cuda")

image = pipe("A magical dragon flying in front of the Himalaya in manga style").images[0]
image

If you want to run the pipeline faster or on different hardware, refer to the optimization docs.

Example Image2Image

# !pip install diffusers==0.3.0
from diffusers import StableDiffusionImg2ImgPipeline
import requests
from PIL import Image
from io import BytesIO

url = "https://scitechdaily.com/images/Dog-Park.jpg"

response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((768, 512))

# using the 115,000 steps checkpoint
pipe = StableDiffusionImg2ImgPipeline.from_pretrained("naclbit/trinart_stable_diffusion_v2", revision="diffusers-115k")
pipe.to("cuda")

images = pipe(prompt="Manga drawing of Brad Pitt", init_image=init_image, strength=0.75, guidance_scale=7.5).images
image

If you want to run the pipeline faster or on different hardware, refer to the optimization docs.

Advanced Usage

Version 2

V2 checkpoint uses dropouts, 10,000 more images, and a new tagging strategy. It's trained longer to improve results while keeping the original aesthetics.

Three flavors

Step 115000/95000 checkpoints were further trained. If you find the style change too significant, you can use the step 60000 checkpoint.

img2img

If you want to run latent-diffusion's stock ddim img2img script with this model, set use_ema to False.

Hardware

• 8xNVIDIA A100 40GB

Training Info

• Custom dataset loader with augmentations: XFlip, center crop, and aspect-ratio locked scaling
• LR: 1.0e-5
• 10% dropouts

Examples

Each image was diffused using K. Crowson's k-lms (from k-diffusion repo) method for 50 steps.

Credits

• Sta, AI Novelist Dev (https://ai-novel.com/) @ Bit192, Inc.
• Stable Diffusion - Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bjorn

📄 License

CreativeML OpenRAIL-M