ComfyUI_LucidNFT
LR-Anchored Multi-Reward Preference Optimization for Generative Real-World Super-Resolution
Nodes (11)
The Two-Node Color Fix That Saves You a Re-Roll
Text, Redux, Control, LoRA, Guidance
The Primer That Tells the Model What 'Bad' Looks Like
A Sampler With Three Real Settings, Because It Doesn't Do CFG
The LoRA merge that turns LucidFlux into LucidNFT
The Flux AE decode — and the wavelet trick that fixes the color shift
The 'super-resolution' that happens before the diffusion model
Conditioning without a CLIP encode — on purpose
The one that actually runs the diffusion
The loader that glues LucidFlux onto your Flux DIT — and the hardest part of this pack
One Dropdown, Twice the Resolution — With a Catch
ComfyUI_LucidNFT
LucidNFT：LR-Anchored Multi-Reward Preference Optimization for Generative Real-World Super-Resolution
Update
- 改成comfyUI的核心封装，提升推理速度/内存占用/changed to comfyUI core wrapper to improve inference speed/memory usage
- 记得给官方LucidNFT点星; Remember to give the official LucidNFT a star
1. 安装
custom_nodes/ 下：
git clone https://github.com/smthemex/ComfyUI_LucidNFT
2. 权重
ComfyUI/models/
├── diffusion_models/flux1-dev-fp8.safetensors # 任意标准 flux DiT（推荐 fp8）
├── vae/UltraFlux.safetensors # 可选：出 2K 时用（F16 /16）
├── lora/LucidNFT_lora_dit.safetensors
├── clip_vision/siglip2-so400m-patch16-512.safetensors # 建议改个名，否则都叫 model.safetensors
└── LucidFlux/ # 与 ComfyUI_LucidFlux 共用同一目录
├── general_swinir_v1.ckpt # DiffBIR 先验
├── lucidflux.pth # 双条件分支（condition_lq/ldr + modulation）
├── prompt_embeddings.pt # 官方预计算文本 embedding（txt 512×4096 + vec 768）
├── lucid_connector.pth # SigLIP→4096 的 Redux 映射（也可直接给 lucidflux.pth）
└── lora_condition/{adapter_config.json, adapter_model.safetensors} # LucidNFT 的条件分支 LoRA（PEFT 原始格式）
转换后（跑一次 tools/convert_dit_lora.py 得到，见第 3 节）：
models/loras/LucidNFT_lora_dit.safetensors # lora_dit 转成的标准 comfy LoRA
下载：lucidflux.pth / prompt_embeddings.pt → W2GenAI/LucidFlux；
lora_condition → W2GenAI/LucidNFT；
lora_dit → smthemex/UltraFlux-v1-gguf（转化后的comfyUI lora）；
SigLIP → google/siglip2-so400m-patch16-512（只下单体模型）；
SwinIR → lxq007/DiffBIR 的 general_swinir_v1.ckpt；
connector → smthem/LucidFLUX-connector。
3 .Example
4. 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}
}
原项目：W2GenAI-Lab/LucidNFT — 觉得有用请给官方点星。