FeMix_HassakuXL: Exploring Blended Anime Styles with FeMix

This repository contains the FeMix_HassakuXL model series, a collection of Stable Diffusion XL models designed for generating high-quality anime-style images. It utilizes a unique blending system (FeMix) to combine different aesthetic styles, offering a range of creative possibilities. This model also uses the CLIP and VAE from HassakuXLv2 to improve the quality

Model Description

The FeMix_HassakuXL series is built on a foundation and enhanced with the FeMix blending system, which combines two distinct styles, Style A and Style B, in various ratios to create models with different characteristics.

Style A: Provides a stable base with great lighting and composition. LoRA friendly.
Style B: Represents a more experimental and stylized approach, potentially with artifacts.

Available Model Variants

This repository includes the following model variants, each representing a different FeMix blend:

FeMix_HassakuXL_A9B1: (90% Style A, 10% Style B) - Provides quality results
FeMix_HassakuXL_A8B2: (80% Style A, 20% Style B) - Provides quality results
FeMix_HassakuXL_A5B5: (50% Style A, 50% Style B) - Great all around for general use
FeMix_HassakuXL_A1B9: (10% Style A, 90% Style B) - Recommended for advanced users who understands how to deal with Style B's artifacts

These model variants also has a variation with CLIP and VAE from HassakuXLv2 to provide better visual quality.

FeMix_HassakuXL_A9B1_v2CLIPVAE: (90% Style A, 10% Style B) - Provides quality results
FeMix_HassakuXL_A8B2_v2CLIPVAE: (80% Style A, 20% Style B) - Provides quality results
FeMix_HassakuXL_A5B5_v2CLIPVAE: (50% Style A, 50% Style B) - Great all around for general use
FeMix_HassakuXL_A1B9_v2CLIPVAE: (10% Style A, 90% Style B) - Recommended for advanced users who understands how to deal with Style B's artifacts

Usage

Recommended Settings

Scheduler: DPM++ 2M Karras, Euler a
Sampling Steps: 20-30
Resolution: SDXL Recommended Resolutions (e.g., 1024x1024, 832x1216)
Clip Skip: 2

Example Prompts

Positive: (masterpiece, best quality, detailed), 1girl, long hair, blue eyes
Negative: (worst quality, bad quality:1.4), (monochrome), greyscale, skin spots, acne, skin blemishes, age spot, watermarks, signature

LoRA Usage

The Style A blends are particularly well-suited for use with LoRA models. Experiment with different LoRAs to further customize the style and details of your generations.

Code Example

from diffusers import StableDiffusionXLPipeline
import torch

model_id = "6chan/FeMix_HassakuXL"
pipe = StableDiffusionXLPipeline.from_pretrained(model_id, torch_dtype=torch.float16, variant="fp16", use_safetensors=True).to("cuda")

prompt = "(masterpiece, best quality, detailed), 1girl, long hair, blue eyes"
negative_prompt = "(worst quality, bad quality:1.4), (monochrome), greyscale, skin spots, acne, skin blemishes, age spot, watermarks, signature"

image = pipe(prompt=prompt, negative_prompt=negative_prompt).images[0]
image.save("generated_image.png")

Limitations

Style B blends (especially A1B9) can produce artifacts and require careful prompting and post-processing.

LoRA effectiveness might vary depending on the blend ratio.

This model is primarily designed for generating anime-style images.

License

This model is released under the Fair AI Public License 1.0-SD. See the LICENSE file for more details.

Acknowledgements

Thanks to the creators of:

Stable Diffusion XL

HassakuXL

Illustrious-XL

6chan
/

FeMix_HassakuXL