ComfyUI-RED-UNO

ComfyUI-RED-UNO

Default 16GB VRAM UNO in context generation ComfyUI-node, using RED-UNO FT model

默认16G显存的UNO ComfyUI in-context生成组件，使用RED-UNO FT模型

RED-UNO FLUX-dev-FT FP8：

https://civitai.com/models/958009/redcraft-or-cads-or-updated-apr14-or-commercial-and-advertising-design-system

image

RED. UNO In-Context (FP8) 4/14/2025

REDAIGC FT Model used to match UNO In-Context Generation

(with improved quality compared to F.1 dev)

解决了FLUX FT底模无法适配UNO组件的问题，FP8权重（显存占用16GB），同时支持Diffusers以及ComfyUI

实测对比DEV生成质量明显提升，对比BF16版本没有明显质量损失，显存占用只有16G，生成时间大概30秒

安装说明

由于ComfyUI-RED-UNO使用Diffusers库进行推理，所以首次使用，会自动下载Diffusers格式的T5和Clip（10G左右）

存放路径是ComfyUI虚拟环境的.cache缓存文件夹，如果是系统级安装的ComfyUI，则会存放至C盘User路径下的.cache（非常占空间）

建议是给ComfyUI的虚拟环境指定一个独立的缓存目录。

FLUX FT底模正常存放在ComfyUI\models\diffusion_models目录下，VAE在ComfyUI\models\vae，UNO-Lora在ComfyUI\models\loras

image

use_fp8和offload 默认勾选，否则24G显存就不够用了，VAE也要使用Diffusers格式的版本否则会报错（ComfyUI目前没有直接支持UNO所以才这么折腾）

VAE版本： https://huggingface.co/GuangyuanSD/16C_vae_Diffusers

---

Diffusers 脚本：

https://github.com/bytedance/UNO

Dit-LoRA 权重：

https://huggingface.co/bytedance-research/UNO

VAE版本：

https://huggingface.co/GuangyuanSD/16C_vae_Diffusers

ComfyUI UNO Nodes

ComfyUI UNO Nodes is a collection of nodes for ComfyUI that allows you to load and use UNO models.

https://github.com/bytedance/UNO comfyui

使用RED-UNO，因为UNO使用的模型格式和ComfyUI不一样，否则就会报错

Missing keys: 236 Unexpected keys: 236

这个错误是因为底模不匹配UNO算法，下载使用这个FLUX底模 https://civitai.com/models/958009 不可以把别的模型改名字替代，因为模型结构不一样！要用Diffusers格式的模型，才可以正确加载。

FP8 support open offload and fp8 support 16GB VRAM

flux model in unet directory and lora in lora directory

clip and t5 will autodownload

UNO0

Paper Analysis

UNO - Less-to-More Generalization Unlocking More Controllability by In-Context Generation

Introduction

Subject-driven image generation, aiming to create new images based on textual descriptions and user-provided reference images, is a central challenge in the field of Artificial Intelligence Generated Content (AIGC). However, existing methods still face significant limitations, particularly regarding data scalability (especially acquiring high-quality multi-subject paired data) and subject expansibility (stably and controllably handling multiple subjects). Researchers from the Intelligent Creation Team at ByteDance address these issues in their paper, "Less-to-More Generalization: Unlocking More Controllability by In-Context Generation," by proposing a novel framework named UNO (Universal aNd cOntrollable), based on an innovative "Less-to-More" generalization paradigm.

Core Challenges and Motivation

Traditional approaches grapple with data scarcity (especially for multi-subject pairs), difficulties in multi-subject control (identity confusion, layout issues), and the inherent trade-off between efficiency and fidelity. UNO's motivation stems from the need to systematically tackle these bottlenecks.

UNO's Core Contributions and Technical Breakdown

UNO presents a comprehensive solution encompassing data, model, and training strategies:

1. Innovative "Model-Data Co-evolution" Paradigm: This is UNO's most groundbreaking conceptual contribution. Instead of passively relying on existing data, it proposes a virtuous cycle where the model itself (even less capable versions) is leveraged to systematically synthesize higher-quality, more complex customized data (evolving from single-subject to multi-subject). This superior data, in turn, trains more powerful ("more controllable") model variants.
2. High-Quality Synthetic Data Automation Pipeline: Utilizes the in-context generation capability of Diffusion Tr