ComfyUI_UltraFlux

UltraFlux:Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios,try it in comfyUI

Update

• 图生图模式上线，加噪伪超分 / i2i is done。
• 因为基于flux ，如果出现人物，推荐使用修手lora，风格lora因为微调图片精度不够，可能会劣化输出，8G显存block number适当从10下调，4G显存，你就从1开始往上测试吧
• Because based on flux, if a character appears, it is recommended to use a hand fix Lora. The style Lora may degrade the output due to insufficient fine-tuning of the image accuracy. The block number of 8GB VRAM should be appropriately reduced from 10, and 4G VRAM should be tested from 1 onwards

1.Installation

In the ./ComfyUI/custom_nodes directory, run the following:

git clone https://github.com/smthemex/ComfyUI_UltraFlux

2.requirements i

• 不装也行，没什么需求

pip install -r requirements.txt

3.Model

• gguf or transformer smthem/UltraFlux-v1-gguf optional/推荐用fp16 ,因为块卸载,只要内存大
• vae rnamae it v1
• diffusers transformer v1 or v1.1 optional/备选 填repo的方式，一般不用
• comfyUI normal T5 and clip-l
• lora, any turbo and style flux lora #任意flux加速和风格lora，部分Lora的精度不够 可能会劣化输出

├── ComfyUI/models/gguf # or transformer
|     ├── UltraFlux-v1-1-BF16.gguf # or Q8
├── ComfyUI/models/diffusion_models # or gguf
|     ├── UltraFlux-v1-1-BF16..safetensors # or e4m3fn
├── ComfyUI/models/vae
|        ├─diffusion_pytorch_model.safetensors  # rename it 换个名字
├── ComfyUI/models/clip
|        ├──t5xxl_fp8_e4m3fn.safetensors
|        ├──clip_l.safetensors 
├── ComfyUI/models/loras 
|        ├──any turbo lora
|        ├──any style lora

4.Example

5.Citation

@misc{ye2025ultrafluxdatamodelcodesignhighquality,
      title={UltraFlux: Data-Model Co-Design for High-quality Native 4K Text-to-Image Generation across Diverse Aspect Ratios}, 
      author={Tian Ye and Song Fei and Lei Zhu},
      year={2025},
      eprint={2511.18050},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.18050}, 
}