English README

<div align="center">

ComfyUI-Wan-latent-Resizer

Learned Latent Resizer for Wan2.1 & Wan2.2 Lightweight · High-fidelity · Unified Upscaling & Downscaling

</div>

---

✨ Overview

A dedicated ComfyUI node that replaces naive interpolation with a trained neural network to resize Wan video latents. A single, lightweight model handles both high-fidelity upscaling and downscaling for Wan2.1 and Wan2.2, supporting arbitrary scale factors.

📰 News

• [2026-05-12] 🎉 Wan2.2 Support Added! A new dedicated model for Wan2.2 is now available, providing unified upscaling and downscaling.
• [2026-05-12] ✨ Unified Model Architecture: Upscaling and downscaling capabilities are now merged into a single, more capable model file for each version.

📸 Comparison Results

Comparison 1 (wan2.1_latent_upscaler_comparison_001.png):

Comparison 2 (wan2.1_latent_upscaler_comparison_002.png):

Comparison 3 (wan2.1_latent_upscaler_comparison_003.mp4):

<video src="examples/wan2.1_latent_upscaler_comparison_003.mp4" controls width="100%"></video>

Left: Latent Interpolation Upscaling (blurry) | Right: Learned Latent Resizing (sharp, our method)

📁 Project Structure

Comfyui-Wan-latent-Resizer/
├── examples/
├── models/
│   ├── wan2.1_latent_resizer_fp16.safetensors
│   └── wan2.2_latent_resizer_fp16.safetensors
├── nodes/
│   ├── __init__.py
│   ├── wan2_1_latent_Resizer_nodes.py
│   └── wan2_2_latent_Resizer_nodes.py
├── workflows/
├── __init__.py
└── README.md

🚀 Key Features

• ✅ Unified Resizing: A single model performs both upscaling and downscaling, simplifying your workflow.
• ✅ Neural Precision: Learned specifically for Wan latents, vastly outperforming traditional interpolation.
• ✅ Arbitrary Scale Factors: Supports continuous scaling across a wide range (e.g., 1.25× – 10×).
• ✅ Broad Compatibility: Seamlessly handles (B,C,T,H,W) video latents and (B,C,H,W) image latents.
• ✅ Minimal Overhead: An ultra-lightweight architecture ensures fast inference with negligible memory cost.
• ✅ Plug-and-Play: Works as a standard ComfyUI node without altering your existing workflow.

🧠 Training Data

• 3,000 high-resolution video clips
• 1,000 high-quality images Diverse content ensures robust performance on a wide range of resizing tasks.

📦 Installation

1. Clone the repository into ComfyUI's custom_nodes folder:

   cd ComfyUI/custom_nodes
   git clone https://github.com/LBH-123-AI/ComfyUI-Wan-latent-Resizer.git
   

2. All required dependencies (torch, einops, safetensors) are already present in a standard ComfyUI environment.

🧩 Usage

Add the "Wan Latent Resizer" node from the menu, select the model version (Wan2.1 or Wan2.2), choose Upscale or Downscale mode, and set your target resolution or scale factor.

Typical workflows:

• Quick Preview: [Wan Latent] → [Wan Latent Resizer] → [Decode]
• High-Quality (Recommended): [Low-res Latent] → [Wan Latent Resizer] → [High-res Refinement] → [Decode] This generates a low-res latent first for speed, upscales it, and then refines it for superior detail, drastically reducing VRAM usage and generation time.

🧪 Model Details

| Model | Parameters | Size | Purpose | | :--- | :--- | :--- | :--- | | wan2.1_latent_resizer_fp16.safetensors | 21.80M | 41.5 MB | Unified resizing for Wan2.1 | | wan2.2_latent_resizer_fp16.safetensors | 51.14M | 97.5 MB | Unified resizing for Wan2.2 |

Architecture: A lightweight ResBlock-based network with scale/target-size conditioning. Input and output are 16-channel Wan latents.

🙏 Acknowledgments

This project is inspired by and builds upon ComfyUi_NNLatentUpscale (https://github.com/Ttl/ComfyUi_NNLatentUpscale). Special thanks to Ttl (https://github.com/Ttl) for the excellent work and open-source contribution.

---

中文版 README

<div align="center">

ComfyUI-Wan-latent-Resizer

Wan2.1 & Wan2.2 专用 Latent 智能缩放节点 轻量 · 高保真 · 放大缩小一体化

</div>

---

✨ 简介

为 Wan 视频/图像生成流程提供的专用 ComfyUI 节点，用训练好的神经网络替代简单的插值算法来完成 latent 缩放。一个轻量级模型即可同时处理 Wan2.1 和 Wan2.2 的高保真放大与缩小任务，并支持任意缩放倍数。

📰 新闻动态

• [2026-05-12] 🎉 新增 Wan2.2 支持！ 现已提供全新的 Wan2.2 专用模型，支持放大与缩小一体化。
• [2026-05-12] ✨ 统一模型架构：每个版本的放大与缩小功能现已合并到一个更强大的模型文件中。

📸 对比效果

对比示例 1（wan2.1_latent_upscaler_comparison_001.png）：

对比示例 2（wan2.1_latent_upscaler_comparison_002.png）：

对比示例 3（wan2.1_latent_upscaler_comparison_003.mp4）：

<video src="examples/wan2.1_latent_upscaler_comparison_003.mp4" controls width="100%"></video>

左：Latent 插值放大（画面模糊） | 右：学习型 Latent 缩放（画面清晰，即本方法）

📁 项目结构

Comfyui-Wan-latent-Resizer/
├── examples/
├── models/
│   ├── wan2.1_latent_resizer_fp16.safetensors
│   └── wan2.2_latent_resizer_fp16.safetensors
├── nodes/
│   ├── __init__.py
│   ├── wan2_1_latent_Resizer_nodes.py
│   └── wan2_2_latent_Resizer_nodes.py
├── workflows/
├── __init__.py
└── README.md

🚀 核心特性

• ✅ 一体式缩放：单模型同时支持放大与缩小，简化工作流。
• ✅ 神经网络高保真：专门针对 Wan latent 学习，效果远超传统插值。
• ✅ 任意缩放倍数：支持连续自定义倍数（如 1.25× – 10×）。
• ✅ 广泛兼容：无缝处理 (B,C,T,H,W) 视频和 (B,C,H,W) 图像 latent。
• ✅ 极低开销：超轻量架构，推理速度极快，资源占用可忽略不计。
• ✅ 即插即用：作为标准 ComfyUI 节点工作，不改变现有流程。

🧠 训练数据

• 3,000 段高清视频素材
• 1,000 张高质量图像素材 多样化的内容确保了模型在各类缩放任务上都有稳健表现。

📦 安装

1. 将仓库克隆到 ComfyUI 的 custom_nodes 文件夹：

   cd ComfyUI/custom_nodes
   git clone https://github.com/LBH-123-AI/ComfyUI-Wan-latent-Resizer.git
   

2. 所有必需的依赖（torch、einops、safetensors）在标准 ComfyUI 环境中已自带。

🧩 使用方法

从节点菜单中添加 "Wan Latent Resizer"，选择模型版本（Wan2.1 或 Wan2.2），选择放大或缩小模式，并设置目标分辨率或缩放倍数即可。

典型流程示例：

• 快速预览：[Wan Latent] → [Wan Latent Resizer] → [解码]
• 高品质输出（推荐）：[低清 Latent] → [Wan Latent Resizer] → [高清重绘] → [解码] 先生成低分辨率 latent 以提高速度，放大后再进行高清细节修复。此流程在保证画质的同时，能显著降低显存占用和生成时间。

🧪 模型详情

| 模型 | 参数量 | 大小 | 用途 | | :--- | :--- | :--- | :--- | | wan2.1_latent_resizer_fp16.safetensors | 21.80M | 41.5 MB | Wan2.1 一体化智能缩放 | | wan2.2_latent_resizer_fp16.safetensors | 51.14M | 97.5 MB | Wan2.2 一体化智能缩放 |

架构：基于残差块的轻量网络，带有缩放倍数/目标尺寸条件输入。输入和输出均为 16 通道 Wan latent。

🙏 致谢

本项目受 ComfyUi_NNLatentUpscale (https://github.com/Ttl/ComfyUi_NNLatentUpscale) 启发并基于其构建，特别感谢原作者 Ttl (https://github.com/Ttl) 的优秀工作和开源贡献。