解码部分推荐用模型，comfyUI的ae解码流程（连标准vae节点时）会有一定的色差，记得给官方LucidNFT点星;
It is recommended to use a model for the decoding part. The ae decoding process in ComfyUI (when connected to the standard VAE node) will have some color differences.Start LucidNFT if you like it.

1.Installation

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

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

2.requirements

pip install -r requirements.txt

3.checkpoints

Any normal flux dit / 任意标准flux模型， KJ的或者官方封装的
Lucid checkpoints links /lucidflux.pth and prompt_embeddings.pt
Lucid lroa links
Siglip512 links / model.safetensors 只下单体模型
DiffBIR links / general_swinir_v1.ckpt
Turbo lora links #optional 可选，8 步起
Flux ae links
Connercter links

├── ComfyUI/models/
|     ├── diffusion_models/any flux dit # 任意flux dit模型 ，就用kj的或者x flux的，名字要带dev 否则跑schnell
|     ├── vae/ae.safetensors #comfy 
|     ├── clip_vision/siglip2-so400m-patch16-512.safetensors  #rename from model.safetensors  最好重命名个，不然都是siglip 的model.safetensors
|     ├── LucidFlux/
|        ├──general_swinir_v1.ckpt
|        ├──lucidflux.pth
|        ├──prompt_embeddings.pt # 已适配，使用时不要连clip
|        ├──lucid_connector.pth # split from lucidflux.pth
|        ├── lora_condition/
|             ├──adapter_config.json
|             ├──adapter_model.safetensors
|        ├── lora_dit/
|             ├──adapter_config.json
|             ├──adapter_model.safetensors

4 .Example

5. Citation

@article{fei2026lucidnft,
  title={LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Generative Real-World Super-Resolution},
  author={Fei, Song and Ye, Tian and Chen, Sixiang and Xing, Zhaohu and Lai, Jianyu and Zhu, Lei},
  journal={arXiv preprint arXiv:2603.05947},
  year={2026}
}
``