spiral@out:~/blog$ cat transmissions/comfyui.md

ComfyUI: Wiring Up AI Models Like a Modular Synth

[2026-10-20] ::self-hosting, ai, image-generation, video, tools

ComfyUI is a node-graph interface and inference engine for generative AI models. You connect boxes with wires to build a pipeline, then run it on your own hardware. It started out as a Stable Diffusion front end and now covers images, video, audio, 3D and text.

It's back on GitHub Trending today, with around 140,000 stars. It's licensed under GPL-3.0.

What it does

Most image-generation tools give you a prompt box and a Generate button. ComfyUI gives you the plumbing instead. A workflow is a graph: one node loads a model, another encodes your prompt, another samples, another decodes the result and saves it. You can rearrange any of it, add a ControlNet or a LoRA, branch off an upscaler, or feed one model's output into another.

The README's model list is long. A few highlights:

  • Images: Stable Diffusion 1.5, SDXL and SD3.5, Flux.1 and Flux.2, Qwen Image, HiDream, Chroma and more, plus editing models such as Flux Kontext and Qwen Image Edit.
  • Video: Wan 2.1 and 2.2, LTX-Video, HunyuanVideo 1.5, CogVideoX and Mochi.
  • Audio: ACE-Step 1.5, Stable Audio 3 and others.
  • 3D and vision: Hunyuan3D 2.1, SAM 3, Depth Anything 3, SUPIR upscaling.
  • Text: Gemma and Qwen3 models, including multimodal inputs.

Built-in tools cover inpainting and outpainting, masks and compositing, model merging, upscaling, frame interpolation, segmentation and depth estimation. It also handles 16-bit PNG, 32-bit EXR and HDR video, which matters if you're doing real post-production work.

How it works

ComfyUI is a Python backend with a browser front end. A few design choices make the node graph more than a pretty diagram:

  • It only runs what changed. Submit the same graph twice and only the first run executes. Change the last node and only that node and its downstream dependencies re-run. Nodes without a complete path to an output don't run at all.
  • Workflows travel with their images. Drag a generated PNG back onto the canvas and ComfyUI rebuilds the entire workflow that made it, seeds included. Workflows also save as plain JSON.
  • It's careful with memory. The README claims it can run "even the biggest open source models on as low as 4GB vram + 8GB ram relatively quickly," using what it calls asynchronous weight streaming, plus model offloading and quantized model support.

Around the core there's a lot more: reusable subgraphs, templates, an "App Mode" that wraps a complicated workflow in a simple UI, a local API for plugging workflows into other software, and a huge ecosystem of custom nodes, managed with the optional ComfyUI-Manager.

Trying it

On Windows and macOS the project recommends the desktop app. There's also a portable Windows build for NVIDIA, AMD and Intel GPUs. The quickest command-line route is comfy-cli:

pip install comfy-cli
comfy install

For a manual install, clone the repo, install the PyTorch build for your GPU (the README has exact commands for NVIDIA CUDA, AMD ROCm, Intel XPU and Apple Silicon), then:

pip install -r requirements.txt
python main.py

Model files go in the matching folders, such as models/checkpoints for checkpoints and models/vae for VAEs. If you already have models downloaded for another UI, extra_model_paths.yaml lets ComfyUI share them instead of keeping duplicates. To add custom nodes, install the manager and launch with it enabled:

pip install -r manager_requirements.txt
python main.py --enable-manager

The fine print

  • Offline by default, with paid extras. Core ComfyUI doesn't download anything unless you ask. It also ships optional "partner nodes" for closed models like Nano Banana and Seedance through Comfy's paid API. Launch with --offline to keep everything local, or --disable-partner-nodes to switch off just those nodes.
  • Custom nodes are other people's code. The ecosystem is ComfyUI's superpower, but every custom node is third-party Python running on your machine. Install the ones you trust.
  • Master moves fast. Stable releases land roughly every two weeks, and the README warns that commits between them "may be very unstable and break many custom nodes." Stick to tagged releases if you depend on a particular node pack.
  • Model licences are separate. GPL-3.0 covers ComfyUI itself. Each model you load comes with its own terms.
  • There's a learning curve. The node graph is the point, but if you just want to type a prompt and get a picture, it's more interface than you need. App Mode and templates soften that.

Who it's for

ComfyUI is for anyone who wants to see and control every step between a prompt and the output: artists building repeatable pipelines, VFX and video people who need high bit depth and compositing, developers who want a local generation backend with an API, and tinkerers who want to try each new open model the week it drops. If a one-box generator has ever felt like a black box, this is the opposite.

The GitHub repo and comfy.org are the places to start, and it's in the resource index now.

End of transmission.

<- BACK TO LOG