ComfyUI
A node-graph engine for composing image, video, audio, 3D and text generation workflows with local models, custom nodes and optional API integrations.
What ComfyUI is useful for
ComfyUI represents generation as a visible node graph. A workflow can connect model loading, text encoding, sampling, decoding and output nodes, then be saved with parameters for later reproduction. The core program, models, custom nodes, templates and external APIs are separate assets with separate permissions and licenses.
Hardware and network boundaries
Resolution, batch size, model count, precision and video length can quickly exceed available VRAM, memory or disk. Core workflows can run offline, while model downloads, Manager, templates, custom-node installers and API nodes may connect to external services. A custom node is executable Python code, not merely a visual theme.
Maintenance note
This page reviews ComfyUI 0.30.0, platform builds, model directories, local workflow reproduction, custom-node permissions, API nodes, offline operation and GPL-3.0 boundaries. Content review date: 2026-08-23.
Save to your cloud drive
Open the cloud drive to get the file directly, or save it for convenient access on another device.
Quark Cloud Drive
RecommendedSave ComfyUI to this cloud drive
Baidu Netdisk
Save ComfyUI to this cloud drive
ComfyUI 0.30.0 installation and workflow safety guide
Start with one trusted model and a small image workflow. Confirm memory use, output paths and workflow saving before adding custom nodes, large video models or API nodes.
Before you start
- Match the Windows NVIDIA, AMD or Intel build, desktop package or manual Python route to the hardware and operating system.
- Prepare models with a clear source, hash, version, base model and license; do not run unknown model or node scripts first.
- Keep separate input, model, output and custom-node directories and reserve space for caches and checkpoints.
Installation steps
- 01
Install the 0.30.0 runtime
Select a supported portable or desktop package, or follow the platform Python and backend instructions, then start with a local-only interface.
- 02
Place one model and load a basic workflow
Put a trusted checkpoint, VAE or text encoder in its expected directory, load a basic template and check for missing or unknown nodes.
- 03
Run a low-resolution test
Use a small size and batch, watch VRAM, memory, logs and the output directory, and save the workflow metadata and seed after a successful run.
Quick start
- 01
Trace the node data flow
Inspect model loading, text encoding, sampling, decoding and saving in order, change one parameter at a time and rerun only the affected part when possible.
- 02
Add custom nodes gradually
Back up the environment, verify each repository and dependency, install one node at a time and remove the newest node first when startup fails.
- 03
Limit API and LAN exposure
Keep API nodes disabled and bind to loopback for local work; remote access needs authentication, a reverse proxy and TLS rather than an open unauthenticated port.
Usage tips
- A workflow is reproducible only when it records the JSON graph, core version, custom-node versions, model hashes, sampler settings and seed.
- Large video, 3D and multi-model graphs need more resources than a small image test; compare one change at a time.
- GPL-3.0 for the core does not automatically cover models, LoRAs, custom nodes, generated material or paid APIs.
Troubleshooting and uninstall
What should I do when VRAM runs out or the result is blank?
Reduce resolution and batch size, close other GPU workloads, try one base workflow and inspect the driver, backend and custom-node logs before adding launch flags.
Why does a workflow say that nodes are missing?
Confirm the workflow source and node repository, install dependencies in a backup environment one by one and restore the last known-good node set instead of installing every unknown package.
- Back up the workflow and model inventoryExport workflow JSON, prompts, seeds, node versions and model hashes, and keep original inputs separate from generated outputs.
- Remove the environment and credentialsStop ComfyUI, remove the portable or virtual environment or uninstall the desktop package, revoke API keys and clean model, node and cache files only when no longer needed.
Frequently asked questions
Does ComfyUI include all models after installation?
No. The core program and model files are separate, and a portable package may contain a runtime without containing the checkpoints required by a workflow.
Can ComfyUI run offline?
Core local workflows can run offline, but model downloads, Manager, templates, custom-node installation and API nodes can add network requests. Disable what the workflow does not need.
Are custom nodes safe because they appear in a node graph?
A custom node usually contains executable Python and may access files, GPU resources or the network. Verify its source and dependencies, and use isolation for important data.