Generate Images and Videos with ComfyUI

Basic idea

ComfyUI is an open-source, node-based web application for AI image and video generation with diffusion models. Instead of a single text box, you connect processing nodes — model loaders, text encoders, samplers, decoders — into a graph that gives full control over every generation step.

• Node-based workflows let you build, modify, and share generation pipelines visually. Workflows save as JSON for versioning, collaboration, and reproducibility.
• Multi-model support covers Stable Diffusion, FLUX for images, Wan 2.1 and HunyuanVideo for video, and NVIDIA Cosmos for world generation.
• GPU-accelerated inference runs on your hardware platform so generation stays local.

What you'll accomplish

Install and run ComfyUI on your hardware platform, then generate images (and optionally video) from a browser UI on port 8188.

Start with the Image Gen Quick Start tab for a lighter host Python install with Z-Image-Turbo. Use the Video Gen Workflow tab for container-based FLUX / Wan / HunyuanVideo / Cosmos / ControlNet workflows.

What to know before starting

Required:

• Basic command-line and terminal usage
• Familiarity with Python virtual environments or Docker containers (depending on install path)

Optional:

• Familiarity with generative AI concepts (prompts, diffusion models, checkpoints)
• Hugging Face account for gated model downloads (container / large-model path)

Supported hardware platforms

Use the matrix below to confirm your hardware platform, OS, memory, and whether multi-node applies.

Prerequisites

Hardware requirements

• Supported hardware platform — see Supported hardware platforms matrix above
• Image Gen Quick Start (host Python): at least ~30 GB free disk; ~24 GB GPU memory for Z-Image-Turbo (12 GB diffusion model plus an 8 GB text encoder, before activations)
• Video Gen Workflow (container + tiered models): at least ~70 GB free disk for Tier 1 (~230 GB for all tiers); peak GPU memory about ~80 GB / ~100 GB / ~120 GB for Tiers 1–3 (see that tab)

Software requirements

• Image Gen Quick Start: Python 3.8+, pip, Git, CUDA toolkit compatible with your GPU, network access to Hugging Face
• Video Gen Workflow: Docker, NVIDIA Container Toolkit, Hugging Face access token, network access to NGC / Hugging Face / GitHub
• Web browser access to port 8188 on the hardware platform

Ancillary files

Playbook assets (Video Gen Workflow) live under assets/ in this repository:

• assets/Dockerfile — Builds the ComfyUI container image from an NGC PyTorch base
• assets/scripts/download-models.sh — Downloads model weights from Hugging Face (hf CLI)
• assets/workflows/*.json — UI workflows for Load in the web UI
• assets/workflow_api/*.api.json — Same graphs in API format for /prompt and automation
• assets/scripts/api_to_ui_workflow.py — Regenerates UI JSON from API JSON if you edit a graph programmatically

For Image Gen Quick Start, clone ComfyUI on GitHub directly (requirements.txt, main.py, and checkpoint directories).

Time & risk

• Estimated time: 2 HOURS (longer on first run when downloading large models)
• Risk level: Medium
  • Model downloads are large and may fail due to network or auth issues
  • Port 8188 must be reachable for the web UI
• Rollback: Remove the virtual environment and clone (Image Gen Quick Start), or stop/remove the container and optionally delete models/ (Video Gen Workflow) — non-destructive to the host OS
• Last Updated: 10/13/2026
  • Video Gen Workflow documents the --disable-dynamic-vram workaround for the HiDream-I1 VBAR allocation error on DGX Station WSL2
  • Image Gen Quick Start now uses the Z-Image-Turbo template