ComfyUI_LucidFlux
ComfyUI_LucidFlux:Caption-Free Universal Image Restoration with a Large-Scale Diffusion Transformer，you can use it in ComfyUI
Nodes (14)
Fixing LucidFlux Color Drift With One Node
The Whole Restoration Brain in One Node (And No, You Don't Need CLIP)
Feed the Diffusion a Pre-Cleaned Image So It Stops Inventing Detail
4K Out of a Latent, If Your Card Can Afford It
CFG Stays at 1, and Why the Sigma Table Cares About Your Resolution
LucidFlux's one-input LoRA merge — mostly the Turbo shortcut, and entirely optional
Turning the restored latent back into pixels — and why 'wavelet' should stay on
The DiffBIR preprocessor that decides what your restoration actually sees
LucidFlux's conditioning stage — and the OOM trap everyone hits first
The custom denoiser that actually runs the restoration (no, a stock KSampler won't do)
This is where a photo becomes a restoration task
PiD pushes the restored latent up to the wall
NVIDIA's PiD decoder, loaded for the 4K leg of LucidFlux
The One-Wire Fix When Your Output Is Half the Size You Asked For
ComfyUI_LucidFlux（comfy 原生 DiT 链版本）
LucidFlux: Caption-Free Universal Image Restoration with a Large-Scale Diffusion Transformer，you can use it in ComfyUI
Update
- 更改封装方式，使用comfyUI的核心复现方法/ change the encapsulation method to use the core method of comfyUI
1.Installation
In the ./ComfyUI/custom_nodes directory, run the following:
git clone https://github.com/smthemex/ComfyUI_LucidFlux
2.checkpoints
2.1 any flux dit / 任意flux模型， KJ的 或者官方封装的
2.2 lucid checkpoints/ links /lucidflux.pth 和prompt_embeddings.pt 和gemma_prompt_embedding.pt
2.3 siglip512 links / model.safetensors 只下单体模型
2.4 DiffBIR links / general_swinir_v1.ckpt
2.5 turbo lora links #optional 可选，8 步起
2.6 lucid_connector.pthlinks
2.7 ultravaelinks ,rename or not
2.8 PIDlinks / PiD_res2kto4k_sr4x_official_flux_distill_4step.pth or PiD_res2k_sr4x_official_flux_distill_4step.pth
2.9 null_caption_embs.ptlinks #PID use
ComfyUI/models/
├── diffusion_models/flux1-dev-fp8.safetensors # 任意 flux-dev（bf16/fp8 均可）
├── PiD_v1pt5_res2kto4k_sr4x_official_flux_distill_4step.pth # v1.5 的 4K 模型（使用PID必需）
├── clip_vision/siglip2-so400m-patch16-512.safetensors
├── vae/UltraFlux.safetensors # 2K 用；1K 可换 ae.safetensors
└── LucidFlux/
├── lucidflux.pth # 双条件分支（必需）
├── lucid_connector.pth # SigLIP→4096 的 redux connector（必需）
├── general_swinir_v1.ckpt # DiffBIR 先验（必需）
├── prompt_embeddings.pt # 官方预计算文本 embedding（**必需**，替代 CLIP/T5）
└── gemma_prompt_embedding.pt / null_caption_embs.pt # 仅 PiD 阶段用
3 Example
- 256 to 4K
4 Citation
LucidFlux
@article{fei2025lucidflux,
title={LucidFlux: Caption-Free Universal Image Restoration via a Large-Scale Diffusion Transformer},
author={Fei, Song and Ye, Tian and Wang, Lujia and Zhu, Lei},
journal={arXiv preprint arXiv:2509.22414},
year={2025}
}