LSRNA

Official code for "Latent Space Super-Resolution for Higher-Resolution Image Generation with Diffusion Models".

Environment (Inference)

conda create -n lsrna python=3.10
conda activate lsrna
pip install -r requirements.txt

Text-to-Image Generation

Note:
Although our LSRNA framework is designed to be compatible with any reference-based method,
this repo provides example code for LSRNA-DemoFusion, as DemoFusion is a pioneering reference-based approach.

CUDA_VISIBLE_DEVICES=0 python main.py \
    --prompt "A well-worn baseball glove and ball sitting on fresh-cut grass." \
    --negative_prompt "blurry, ugly, duplicate, poorly drawn, deformed, mosaic" \
    --height 2048 \
    --width 2048 \
    --seed 0 \
    --lsr_path "lsr/swinir-liif-latent-sdxl.pth" \
    --rna_min_std 0.0 \
    --rna_max_std 1.2 \
    --inversion_depth 30 \
    --save_dir "results" \
    #--low_vram

Feel free to adjust the RNA hyperparameters (e.g., --rna_max_std) to adjust the level of detail in the generated images. If you’re running out of VRAM, enable the low-VRAM mode with --low_vram. We also provide a run.sh script for the generation.

Visual Comparison

Additional results can be found on the project page.

Citation

@inproceedings{jeong2025latent,
  title={Latent space super-resolution for higher-resolution image generation with diffusion models},
  author={Jeong, Jinho and Han, Sangmin and Kim, Jinwoo and Kim, Seon Joo},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={2355--2365},
  year={2025}
}

Acknowledgement

This repo is based on DemoFusion and LIIF.