๐ Mobius: Redefining State-of-the-Art in Debiased Diffusion Models
Mobius is a diffusion model that breaks the boundaries of domain-agnostic debiasing and representation realignment, offering unparalleled generalization across various styles and domains.
๐ Quick Start
The following is a quick start guide to using Mobius with ๐งจ diffusers:
import torch
from diffusers import (
StableDiffusionXLPipeline,
KDPM2AncestralDiscreteScheduler,
AutoencoderKL
)
vae = AutoencoderKL.from_pretrained(
"madebyollin/sdxl-vae-fp16-fix",
torch_dtype=torch.float16
)
pipe = StableDiffusionXLPipeline.from_pretrained(
"Corcelio/mobius",
vae=vae,
torch_dtype=torch.float16
)
pipe.scheduler = KDPM2AncestralDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.to('cuda')
prompt = "mystery"
negative_prompt = ""
image = pipe(
prompt,
negative_prompt=negative_prompt,
width=1024,
height=1024,
guidance_scale=7,
num_inference_steps=50,
clip_skip=3
).images[0]
image.save("generated_image.png")
โจ Features
Domain-Agnostic Debiasing: A Groundbreaking Approach
Domain-agnostic debiasing is a novel technique pioneered by Corcel. This innovative approach aims to remove biases inherent in diffusion models without limiting their ability to generalize across diverse domains. Traditional debiasing methods often focus on specific domains or styles, resulting in models that struggle to adapt to new or unseen contexts. In contrast, domain-agnostic debiasing ensures that the model remains unbiased while maintaining its versatility and adaptability.
Surpassing the State-of-the-Art
Mobius outperforms existing state-of-the-art diffusion models in several key areas:
- Unbiased generation: Mobius generates images that are virtually free from the inherent biases commonly found in other diffusion models, setting a new benchmark for fairness and impartiality across all domains.
- Exceptional generalization: With its unparalleled ability to adapt to an extensive range of styles and domains, Mobius consistently delivers top-quality results, surpassing the limitations of previous models.
- Efficient fine-tuning: The Mobius base model serves as a superior foundation for creating specialized models tailored to specific tasks or domains, requiring significantly less fine-tuning and computational resources compared to other state-of-the-art models.
๐ก Usage Tips
- CFG between 3.5 and 7:
- Use 3.5 for extreme realism and skin detailing.
- Use 7 for artistic, anime, surrealism, and so on.
- Requires a CLIP skip of -3
- Sampler: DPM++ 3M SDE
- Scheduler: Karras
- Steps: 50
- Resolution: 1024x1024
Please also consider using these keep words to improve your prompts: best quality, HD, '~~aesthetic~~'.
๐ License
This project is licensed under the Apache-2.0 license.
๐ฅ Credits
Made by Corcel
๐ Widget Examples
Input Text |
Output Image |
movie scene screencap, cinematic footage. thanos smelling a little yellow rose. extreme wide angle, |
1man.png |
god |
god.png |
A tiny robot taking a break under a tree in the garden |
robot.png |
mystery |
mystery.png |
a cat wearing sunglasses in the summer |
cat.png |
robot holding a sign that says โa storm is comingโ |
storm.png |
The Exegenesis of the soul, captured within a boundless well of starlight, pulsating and vibrating wisps, chiaroscuro, humming transformer |
soul.png |
anime boy, protagonist, best quality |
animeboy.png |
natural photography of a man, glasses, cinematic, |
glasses.png |
if I could turn back time |
time.png |
("Mobius" text logo) powerful aura, swirling power, cinematic |
mobius.png |
the backrooms |
backrooms.png |