Awesome

UltraPixel: Advancing Ultra-High-Resolution Image Synthesis to New Peaks (NeurIPS 2024)

UltraPixel is designed to create exceptionally high-quality, detail-rich images at various resolutions, pushing the boundaries of ultra-high-resolution image synthesis. For more details and to see more stunning images, please visit the Project Page. The arXiv version of the paper contains compressed images, while the full paper features uncompressed, high-quality images.

🔥 Updates:

• 2024/09/26: 🎉 UltraPixel has been accepted to NeurIPS 2024!
• 2024/09/19: 🤗 We released the HuggingFace Space, thanks to the HF team and Gradio! Gradio interface for text-to-image inference is also provided, and please see Inference section!
• 2024/09/19: We have updated the versions of PyTorch and Torchvision in our environment. On an RTX 4090 GPU, generating a 2560×5120 image (without stage_a_tiled) now takes approximately 60 seconds, compared to about three minutes in the previous setup.

Getting Started

1. Install dependency by running:

pip install -r requirements.txt

2. Download pre-trained models from StableCascade model downloading instructions. Small-big models (the small model for stage b and the big model for stage with bfloat16 format are used.) The big-big setting is also supported, while small-big favors more efficiency.

3. Download newly added parameters of UltraPixel from here.

Note: All model downloading urls are provided here. They should be put in the directory models.

Inference

Text-guided Image Generation

We provide Gradio interface for inference. Run by :

CUDA_VISIBLE_DEVICES=0 python app.py

Or generate an image by running:

CUDA_VISIBLE_DEVICES=0 python inference/test_t2i.py

Tips: To generate aesthetic images, use detailed prompts with specific descriptions. It's recommended to include elements such as the subject, background, colors, lighting, and mood, and enhance your prompts with high-quality modifiers like "high quality", "rich detail", "8k", "photo-realistic", "cinematic", and "perfection". For example, use "A breathtaking sunset over a serene mountain range, with vibrant orange and purple hues in the sky, high quality, rich detail, 8k, photo-realistic, cinematic lighting, perfection". Be concise but detailed, specific and clear, and experiment with different word combinations for the best results.

Several example prompts are provided here.

It is recommended to add "--stage_a_tiled" for decoding in stage a to save memory.

The table below show memory requirements and running times on different GPUs. For the A100 with 80GB memory, tiled decoding is not necessary.

On 80G A100:

On 32G V100 (only works using float32 on Stages C and B):

On 24G RTX4090:

Personalized Image Generation

The repo provides a personalized model of a cat. Download the personalized model here and run the following command to generate personalized results. Note that in the text command you need to use identifier "cat [roubaobao]" to indicate the cat.

CUDA_VISIBLE_DEVICES=0 python inference/test_personalized.py

Controlnet Image Generation

Download Canny ControlNet provided by StableCascade and run the command:

CUDA_VISIBLE_DEVICES=0 python inference/test_controlnet.py

Note that ControlNet is used without further fine-tuning, so the supported highest resolution is 4K, e.g., 3840 * 2160, 2048 * 2048.

T2I Training

Put all your images and captions into a folder. Here's an example training dataset here for reference. Start training by running:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train/train_t2i.py configs/training/t2i.yaml

Personalized Training

Put all your images into a folder. Here's an expample training dataset here. The training prompt can be described as: a photo of a cat [roubaobao].

Start training by running:

CUDA_VISIBLE_DEVICES=0,1 python train/train_personalized.py \
configs/training/lora_personalization.yaml

Citation

@article{ren2024ultrapixel,
  title={UltraPixel: Advancing Ultra-High-Resolution Image Synthesis to New Peaks},
  author={Ren, Jingjing and Li, Wenbo and Chen, Haoyu and Pei, Renjing and Shao, Bin and Guo, Yong and Peng, Long and Song, Fenglong and Zhu, Lei},
  journal={arXiv preprint arXiv:2407.02158},
  year={2024}
}

Contact Information

To reach out to the paper’s authors, please refer to the contact information provided on the project page.

Acknowledgements

This project is build upon StableCascade and Trans-inr. Thanks for their code sharing ：）