---
license: cc-by-4.0
---
![Intro Image](cosmicman_samples.png)

CosmicMan is a text-to-image foundation model specialized for generating high-fidelity human images. For more information, please refer to our research paper: [CosmicMan: A Text-to-Image Foundation Model for Humans](https://arxiv.org/abs/2404.01294). Our model is based on [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0). This repository provide UNet checkpoints for CosmicMan-SDXL. 


## Requirements

```python
conda create -n cosmicman python=3.10
source activate cosmicman
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install accelerate diffusers datasets transformers botocore invisible-watermark bitsandbytes gradio==3.48.0
```


### Quick start with [Gradio](https://www.gradio.app/guides/quickstart)

To get started, first install the required dependencies, then run:

```
python demo_sdxl.py
```
Let's have a look at a simple example using the `http://your-server-ip:port`.


## Inference

```python
import torch
from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, UNet2DConditionModel, EulerDiscreteScheduler
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

base_path = "stabilityai/stable-diffusion-xl-base-1.0"
refiner_path = "stabilityai/stable-diffusion-xl-refiner-1.0"
unet_path = "cosmicman/CosmicMan-SDXL"

# Load model.
unet = UNet2DConditionModel.from_pretrained(unet_path, torch_dtype=torch.float16)
pipe = StableDiffusionXLPipeline.from_pretrained(base_path, unet=unet, torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.scheduler = EulerDiscreteScheduler.from_pretrained(base_path, subfolder="scheduler", torch_dtype=torch.float16)

refiner = StableDiffusionXLImg2ImgPipeline.from_pretrained(base_path,torch_dtype=torch.float16, use_safetensors=True).to("cuda") # using "base_path" instead of "refiner_path" may have a better result

# Generate image.
positive_prompt = "A fit Caucasian elderly woman, her wavy white hair above shoulders, wears a pink floral cotton long-sleeve shirt and a cotton hat against a natural landscape in an upper body shot"
negative_prompt = ""
image = pipe(positive_prompt, num_inference_steps=30, 
        guidance_scale=7.5, height=1024, 
        width=1024, negative_prompt=negative_prompt, output_type="latent").images[0]
image = refiner(positive_prompt, negative_prompt=negative_prompt, image=image[None, :]).images[0].save("output.png")
```


## Citation Information
```
@article{li2024cosmicman,
  title={CosmicMan: A Text-to-Image Foundation Model for Humans},
  author={Li, Shikai and Fu, Jianglin and Liu, Kaiyuan and Wang, Wentao and Lin, Kwan-Yee and Wu, Wayne},
  journal={arXiv preprint arXiv:2404.01294},
  year={2024}
}

```