metadata

license: openrail++
language:
  - en
thumbnail: >-
  https://huggingface.co/Linaqruf/hermitage-xl/resolve/main/sample_images/thumbnail.png
pipeline_tag: text-to-image
tags:
  - stable-diffusion
  - stable-diffusion-diffusers
inference: true
widget:
  - text: >-
      masterpiece, best quality, 1girl, brown hair, green eyes, colorful,
      autumn, cumulonimbus clouds, lighting, blue sky, falling leaves, garden
    example_title: example 1girl
  - text: >-
      masterpiece, best quality, 1boy, medium hair, blonde hair, blue eyes,
      bishounen, colorful, autumn, cumulonimbus clouds, lighting, blue sky,
      falling leaves, garden
    example_title: example 1boy
library_name: diffusers

Hermitage XL

Overview

Hermitage XL is a high-resolution, latent text-to-image diffusion model. The model has been fine-tuned using a learning rate of 4e-7 over 5000 steps with a batch size of 16 on a curated dataset of superior-quality anime-style images. This model is derived from Stable Diffusion XL 1.0.

e.g. 1girl, white hair, golden eyes, beautiful eyes, detail, flower meadow, cumulonimbus clouds, lighting, detailed sky, garden

Use it with the Stable Diffusion Webui
Use it with 🧨 diffusers
Use it with the ComfyUI

Features

High-Resolution Images: The model trained with 1024x1024 resolution. The model is trained using NovelAI Aspect Ratio Bucketing Tool so that it can be trained at non-square resolutions.
Anime-styled Generation: Based on given text prompts, the model can create high quality anime-styled images.
Fine-Tuned Diffusion Process: The model utilizes a fine-tuned diffusion process to ensure high quality and unique image output.

Model Details

Developed by: Linaqruf
Model type: Diffusion-based text-to-image generative model
Model Description: This is a model that can be used to generate and modify anime-themed images based on text prompts.
License: CreativeML Open RAIL++-M License
Finetuned from model: Stable Diffusion XL 1.0

How to Use:

Download Hermitage XL here, the model is in .safetensors format.
You need to use Danbooru-style tag as prompt instead of natural language, otherwise you will get realistic result instead of anime
You can use any generic negative prompt or use the following suggested negative prompt to guide the model towards high aesthetic generationse:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

And, the following should also be prepended to prompts to get high aesthetic results:

masterpiece, best quality, illustration, beautiful detailed, finely detailed, dramatic light, intricate details

🧨 Diffusers

Make sure to upgrade diffusers to >= 0.18.2:

pip install diffusers --upgrade

In addition make sure to install transformers, safetensors, accelerate as well as the invisible watermark:

pip install invisible_watermark transformers accelerate safetensors

Running the pipeline (if you don't swap the scheduler it will run with the default EulerDiscreteScheduler in this example we are swapping it to EulerAncestralDiscreteScheduler:

import torch
from torch import autocast
from diffusers.models import AutoencoderKL
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler

model = "Linaqruf/hermitage-xl"
vae = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae")

pipe = StableDiffusionXLPipeline.from_pretrained(
    model, 
    torch_dtype=torch.float16, 
    use_safetensors=True, 
    variant="fp16",
    vae=vae
    )

pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.to('cuda')

prompt = "masterpiece, best quality, 1girl, green hair, sweater, looking at viewer, upper body, beanie, outdoors, watercolor, night, turtleneck"
negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"

image = pipe(
    prompt, 
    negative_prompt=negative_prompt, 
    width=1024,
    height=1024,
    guidance_scale=12,
    target_size=(1024,1024),
    original_size=(4096,4096),
    num_inference_steps=50
    ).images[0]

image.save("anime_girl.png")

Limitation

This model inherit Stable Diffusion XL 1.0 limitation
This model is overfitted and cannot follow prompts well, because it's fine-tuned for 5000 steps with small scale datasets.
It's only a preview model to find good hyperparameter and training config for Stable Diffusion XL 1.0

Example

Here is some cherrypicked samples and comparison between available models: