Patreon
Home / Guides / SwarmUI Install Guide: ComfyUI Backend, Presets and the Mode

SwarmUI Install Guide: ComfyUI Backend, Presets and the Model Downloader

By Furkan Gözükara · Updated Nov 21, 2026

This SwarmUI install guide follows the route SECourses uses on Windows: install the ComfyUI backend first, install SwarmUI with one double-click, connect the two, import the Amazing SwarmUI Presets and download every model a preset needs with the model downloader. The SwarmUI installer, the presets and the downloader come in one zip that is available to members on Patreon; the current V195 ships 105 presets. The same installers run on RunPod, SimplePod, Massed Compute and Linux.

What you need

PartWhere it comes fromNotes
Windows requirements1-click requirements installer, once per PCIncludes the .NET 10 SDK that SwarmUI needs
ComfyUI backendComfyUI installer postMandatory for SwarmUI; Torch 2.13 and CUDA 13
SwarmUI installer, downloader and presetsSwarmUI installer postSwarmUI_Model_Downloader_v195.zip

The requirements step is covered in the AI requirements installer guide, and the backend in the ComfyUI 1-click installer guide. The SwarmUI post is clear that you should use this ComfyUI backend installer, because some of the newest features may fail otherwise.

Step-by-step install on Windows

The SwarmUI full tutorial shows this route on a fresh Windows PC with an RTX 5090.

1. Install the ComfyUI backend

Create a folder for each package, extract both zips and run Windows_Install_Or_Update_ComfyUI.bat from the ComfyUI package first. SwarmUI generates through this ComfyUI, so it has to be ready before you connect it.

2. Install SwarmUI

Run Windows_Install_or_Update_SwarmUI.bat from the SwarmUI package. It installs SwarmUI together with the SECourses extensions (the tutorial counts 11) and, if SwarmUI cannot be installed, it tells you why and what to do next. In the first-run setup, the Windows backend tutorial calls choosing None for the ComfyUI backend a crucial step, because you add your own backend afterwards.

3. Connect the ComfyUI backend

  1. Open Server → Backends and add a ComfyUI Self-Starting backend.
  2. Point it at the main.py of your ComfyUI, for example E:\Comfy_UI_V105\ComfyUI\main.py (on a cloud pod, /workspace/ComfyUI/main.py).
  3. Add the recommended arguments: --disable-pinned-memory --use-sage-attention --enable-triton-backend --disable-comfy-compiler.
  4. For a second GPU, add another backend with GPU ID 1. SwarmUI then spreads generations across both cards.

Sage Attention is optional. If a model gives black output with it, enable Display Advanced Options, open Advanced Sampling and set Preferred DType to Default (16 bit). For the highest quality, the post names --use-pytorch-cross-attention as the attention argument.

4. Import the presets

Run Windows_Preset_Delete_Import.bat. It backs up your current presets, then replaces all of them, your own included, with the latest file (Amazing_SwarmUI_Presets_v92.json in V195).

To keep your own presets, click Import in the Presets tab instead, choose the latest preset file and turn on Overwrite existing presets. Presets with the same names are updated and your differently named ones stay. To change one of ours, duplicate and rename it first, so later updates never overwrite it.

Use presets with one routine: Quick Tools → Reset Params to Default, then Direct Apply the preset. Filter the Presets tab by model name or by date code, for example "LongCat" or "261117".

5. Download models with the model downloader

Start Windows_Start_Download_Models_App.bat and check the download path. With SwarmUI, download into the SwarmUI Models folder. For ComfyUI only, paste your ComfyUI models folder, select the ComfyUI folder structure and click Remember Settings. The downloader also serves Automatic1111 and Forge Web UI (Classic and Neo).

  • Bundles: one ComfyUI and SwarmUI Bundles tab (V195) holds every core bundle, so one click gets all the files a preset family needs. The LongCat-Video-Avatar 1.6 Core Bundle, for example, has 11 files.
  • Formats: many models come as INT8 ConvRot in the core bundles, W4A8 in Low VRAM bundles and BF16 in original precision; GGUF and NVFP4 files are listed too.
  • No repeat downloads: files you already have are verified, not downloaded again. Since V184 the downloader also keeps files whose header SwarmUI rewrote, after checking that every tensor byte matches. In one test the Fibo 2.0 Core Bundle was checked and kept in 25 seconds instead of downloading 20.50 GB again.
  • Any other model: paste a Civitai or Hugging Face link into the URL downloader.
  • Which bundle for which preset: Which_Bundles_Downloads_Which_Preset_Models.html lists the models of every preset.

The download queue shows how many files finished, failed or were cancelled. In v187 the catalog held 36 bundles and 534 models in 13 sections.

6. Generate your first image or video

Refresh the model list, reset the parameters, apply a preset and click Generate; every result waits in History, and Reuse Parameters loads its settings again. The SwarmUI beginner's guide to AI video, written by SwarmUI developer mcmonkey, adds two tips: start with base models and simple prompts, and for direct image to video set Init Image Creativity to 0, a common mistake.

Update in the right order

SwarmUI runs on your ComfyUI backend, so the update video updates ComfyUI first. Always extract the new zips over the old folders and overwrite every file:

  1. ComfyUI: extract the latest ComfyUI zip into your ComfyUI installer folder, close SwarmUI and run Windows_Install_Or_Update_ComfyUI.bat.
  2. Models: extract the latest SwarmUI_Model_Downloader zip over your downloader folder, start the downloader and get any new core bundle.
  3. SwarmUI: run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI and the extensions, stages each update, rolls a failed one back and verifies every extension against the new build.
  4. Presets: run Windows_Preset_Delete_Import.bat and refresh the Presets tab.

If you use SwarmUI's own ComfyUI instead, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start.

Use SwarmUI presets in ComfyUI and share one models folder

SwarmUI can export any preset as a ComfyUI workflow in one click. The SwarmUI to ComfyUI tutorial runs a Z Image Turbo ControlNet preset and a Wan 2.2 image-to-video preset this way, after installing the SwarmUI Extra Nodes in ComfyUI. It also points ComfyUI to the SwarmUI model folders through extra_model_paths.yaml, so both apps share one folder instead of keeping duplicate checkpoints and LoRAs.

FAQ

Do I need to install ComfyUI separately for SwarmUI?

Yes. The SwarmUI post lists the SECourses ComfyUI backend as mandatory, and the 2025 Windows tutorial explains why a manually installed backend is better for SwarmUI performance. Install it first, then connect its main.py in Server → Backends.

Will importing the presets delete my own presets?

Windows_Preset_Delete_Import.bat backs up your presets and then replaces all of them. To keep yours, import the preset file from the Presets tab with Overwrite existing presets turned on.

Why do I get black images in SwarmUI?

With Sage Attention, a few models can give black output. Set Preferred DType to Default (16 bit) under Advanced Sampling, or start the backend with --use-pytorch-cross-attention.

Does the downloader download models I already have again?

No. Existing files are verified and skipped. Since V184 it also keeps files whose header SwarmUI changed, as long as the tensor data matches the Hugging Face file.

Can I run SwarmUI on cloud GPUs?

Yes. The same installers run on RunPod, SimplePod and Massed Compute; the cloud setup video shows RunPod and Massed Compute with multi-GPU backends. See also RunPod vs Massed Compute vs SimplePod.

Related