ComfyUI
A node-graph app and local server for running open diffusion models (image, video, audio, 3D) as reusable workflows, aimed at creators and pipeline builders.
Comfy Org often repackages weights for ComfyUI (Hugging Face org Comfy-Org); the ids below are the repos the official ComfyUI tutorials download from.
docs.comfy.org also has Partner-node tutorials for hosted APIs (e.g. LTX-2.5, MiniMax H3, Ideogram v4); only native local workflows are counted here.
The repo moved from comfyanonymous/ComfyUI to Comfy-Org/ComfyUI (GitHub redirects).
At a glance
- Built by
- Comfy Org
- Status
- Active
- Latest release
- v0.42.1 (2026-10-19)
- License
- GPL-3.0
- Written in
- Python
- Serves
- Image, Video, Audio
- Hardware
- NVIDIA GPU, AMD GPU, Intel GPU, Apple Silicon, Ascend NPU, Cambricon MLU, Iluvatar Corex, CPU
- API
- Local HTTP + WebSocket API (POST a workflow JSON to /prompt; progress on /ws)
- Quantization
- FP8 (scaled), NVFP4, INT8
From ComfyUI's own repo and docs, checked 21 Oct 2026.
Get started
pip install comfy-cli && comfy installpython main.pyWhat it's good at
- Workflows, not one model per call
Models, text encoders, VAEs, LoRAs and ControlNets are wired together as a node graph saved as JSON; the same graph can be queued over the local API. source ↗
- Built for consumer hardware
Smart VRAM/RAM management, model offloading and asynchronous weight streaming; the README claims the biggest open models run on as little as 4 GB VRAM + 8 GB RAM. source ↗
- Fast, broad model support
Native support for SD3.5, FLUX.2, Qwen Image, Ideogram 4, Krea 2, LTX-2/2.3, MiniMax H3, Stable Audio 3 and many more, usually with ready-to-run templates at launch. source ↗
- Frequent releases
A weekly release cycle with a new stable core version roughly every two weeks; commits between tags may break custom nodes. source ↗
- Offline by default, paid nodes optional
Core downloads nothing unless asked; optional paid Partner/API nodes call closed models and can be disabled with --offline or --disable-partner-nodes. source ↗
Catalog models it runs 12
- Comfy-Org/stable-audio-3
Text-to-audio via a dedicated subgraph node, with optional Qwen prompt expansion.
- Comfy-Org/Qwen-Image-2.1
Native workflow for generation and editing, including alpha-channel output.
- black-forest-labs/flux-3-image-klein-comfy
Templates need ComfyUI 0.41.0 or later (docs say it shipped in the 0.41.1 backport); gated BFL repo with int8, bf16 and nvfp4 checkpoints.
- Comfy-Org/flux2-dev
Native workflow; BFL's model card also says FLUX.2 [dev] is available in ComfyUI.
- Comfy-Org/flux2-klein-4B
Text-to-image and image-editing workflows; BFL's model card also says it is available in ComfyUI.
- Comfy-Org/Ideogram-4
Native workflow using Comfy-Org's fp8-scaled repack of the open weights.
- Comfy-Org/Krea-2
Both RAW (base) and Turbo (distilled) variants.
- stabilityai/stable-diffusion-3.5-large
Stability's model card recommends ComfyUI for local use and ships an example workflow; ComfyUI's SD3 examples cover sd3.5_large.
- stabilityai/stable-diffusion-3.5-medium
Stability's model card recommends ComfyUI for local use and ships an example workflow; ComfyUI's SD3 examples cover sd3.5_medium.
- Lightricks/LTX-2.3
Native text-to-video, image-to-video, FLF2V and audio-driven workflows; Lightricks' card recommends the LTXVideo nodes in ComfyUI.
- Lightricks/LTX-2.5
Native workflows with synchronized audio; Lightricks says official LTX-2.5 workflow templates ship in ComfyUI.
- Comfy-Org/MiniMax-H3
Native text-, image- and reference-to-video workflows; MiniMax's model card links the Comfy tutorial and templates.
Only models with an official source are listed. ComfyUI may run others with the same architecture; see its supported-models page.