ComfyUI-HiDiffusion

ComfyUI-HiDiffusion is a free set of custom nodes that brings the HiDiffusion technique to ComfyUI, letting Stable Diffusion and SDXL models generate images at resolutions far beyond their native training size. The project has more than 140 stars on GitHub and is released under the GPL-3.0 license.

Overview

HiDiffusion, developed by Megvii Research, restructures the U-Net inside diffusion models — with ideas such as resolution-adjusted attention and window attention — so that a model trained at modest resolutions can produce coherent pictures at several times that size, instead of duplicating objects or smearing details. This extension wraps that method in ComfyUI nodes. A dedicated SDXL node exposes the controls artists actually tune: checkpoint name, toggles for RAU-Net and window attention, optimization switches, steps, guidance scale and scheduler. The author demonstrates sample output at 4096×2304 from an SDXL checkpoint, and links to the original HiDiffusion project for the underlying method and specifications. The extension is candidly marked as under development and testing, and the author recommends trying it in a separate ComfyUI installation first — honest guidance worth following before adding it to a production workflow. Its most recent updates landed in 2025.

Key Features

• High-resolution generation – push SD and SDXL models to outputs such as 4096×2304 without the usual artifacts.
• Dedicated SDXL node – checkpoint, RAU-Net and window-attention toggles, steps, guidance scale and scheduler controls.
• Resolution-adjusted U-Net – implements HiDiffusion’s modified attention to keep very large images structurally coherent.
• Native ComfyUI workflow – drops into existing node graphs alongside your usual samplers and loaders.
• Documented samples – example images published with full settings (model, steps, CFG, scheduler) to reproduce results.
• Free and open source – GPL-3.0, maintained by an independent ComfyUI community developer.

💳 Pricing

ComfyUI-HiDiffusion is completely free. It is an open-source extension for the equally free ComfyUI; your only costs are your own GPU time and electricity.

Who Is It For

It is for ComfyUI artists and Stable Diffusion power users who need poster-size or widescreen outputs from SDXL checkpoints, and who want more control over high-resolution artifacts than a simple upscale pass provides.