[2025/06/03] We release the updated checkpoint flux_wavelet_v2_sc using SC-VAE and Aesthetic-Train-V2.
Please refer to our new arxiv paper Ultra-High-Resolution Image Synthesis: Data, Method and Evaluation for more details.
The training code of SC-VAE is released at sc-vae.
The Aesthetic-Train-V2 training set is released at huggingface/Aesthetic-Train-V2.
We present Diffusion-4K, a novel framework for direct ultra-high-resolution image synthesis using text-to-image diffusion models. The core advancements include:
(1) Aesthetic-4K Benchmark: addressing the absence of a publicly available 4K image synthesis dataset, we construct Aesthetic-4K, a comprehensive benchmark for ultra-high-resolution image generation. We curated a high-quality 4K dataset with carefully selected images and captions generated by GPT-4o. Additionally, we introduce GLCM Score and Compression Ratio metrics to evaluate fine details, combined with holistic measures such as FID, Aesthetics and CLIPScore for a comprehensive assessment of ultra-high-resolution images.
(2) Wavelet-based Fine-tuning: we propose a wavelet-based fine-tuning approach for direct training with photorealistic 4K images, applicable to various latent diffusion models, demonstrating its effectiveness in synthesizing highly detailed 4K images.
Consequently, Diffusion-4K achieves impressive performance in high-quality image synthesis and text prompt adherence, especially when powered by modern large-scale diffusion models (e.g., SD3-2B and Flux-12B).
- Aesthetic-4K dataset is released at huggingface/Aesthetic-4K.
Install the required packages:
pip install -r requirements.txt
Download the dataset in Aesthetic-4K folder, and the pretrained latent diffusion models(SD3-2B and Flux-12B) in pretrain folder.
Modify the configuration in ds_config.yaml accordingly (e.g. num_processes), and then run the script for training Flux-12B:
sh train_flux.sh
Flux-12B-WLF checkpoint is released at flux_wavelet.
Generate images with Flux-12B:
CUDA_VISIBLE_DEVICES=0 python test_flux.py --guidance_scale 5.0 --height 3072 --width 4096 --seed 0 --num_inference_steps 50 --checkpoint_path $CHECKPOINT_PATH --prompt $PROMPT
For 4K image generation, it is recommended to generate at the resolution of (4096, 3072) or (3072, 4096) with guidance scale of 5 or 5.5.
Generate images for evaluation and compute FID, CLIPScore and Aesthetics:
python generate_and_eval_fid.py --size 2048 --checkpoint_path $CHECKPOINT_PATH --save_path $SAVE_PATH
Compute GLCM Score and Compression Ratio with generated images:
python eval.py --size 2048 --img_root $SAVE_PATH
We are grateful for the following awesome projects and models when implementing Diffusion-4K:
If you find our paper or dataset is helpful in your research or applications, generously cite with:
@inproceedings{zhang2025diffusion4k,
title={Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models},
author={Zhang, Jinjin and Huang, Qiuyu and Liu, Junjie and Guo, Xiefan and Huang, Di},
year={2025},
booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
}
@misc{zhang2025ultrahighresolutionimagesynthesis,
title={Ultra-High-Resolution Image Synthesis: Data, Method and Evaluation},
author={Zhang, Jinjin and Huang, Qiuyu and Liu, Junjie and Guo, Xiefan and Huang, Di},
year={2025},
note={arXiv:2506.01331},
}