ComfyUI is the standard way to run open-weight AI video models on your own hardware. It is a free, node-based interface where you wire together models, prompts, and samplers into reusable workflows — no code required after setup. In 2026 the ecosystem is mature: first-party templates exist for LTX-2.5 and Wan 2.2, and a range of quantized community builds make local video generation possible on GPUs from 8 GB upward.
This guide walks you through installing ComfyUI, picking a video model your GPU can actually run, and building your first three video workflows.
What you need before you start
ComfyUI runs on Windows, Linux, and macOS, with support for NVIDIA, AMD, Intel, and Apple Silicon GPUs. You need Python 3.10 or newer and enough disk space for model weights — video checkpoints are much heavier than image models, typically 5 to 40 GB each.
The single most important constraint is VRAM. Realistic guidance from model developers for 2026:
- 8–12 GB VRAM: quantized builds of Wan 2.2 TI2V-5B, or FramePack for image-to-video
- 14–24 GB VRAM: HunyuanVideo 1.5 with model offloading, or distilled FP8 LTX builds
- 32 GB VRAM: LTX-2.5 at its official minimum
- 48 GB+ VRAM: Wan 2.2 A14B at full quality
Do not trust random download pages claiming open weights for "Wan 2.5", "Wan 2.6", or "Wan 2.7". As of late 2026, Alibaba's official open weights stop at Wan 2.2; later versions are API-only. Anything advertising a Wan 2.7 checkpoint is not what it claims to be. Download weights only from the official Hugging Face organizations (Wan-AI, Lightricks, Tencent-HunyuanVideo).
Step 1: Install ComfyUI
Clone the official repository and install dependencies:
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
python main.py
Then open http://127.0.0.1:8188 in your browser. This is your workspace from now on — everything happens in this one tab.
Step 2: Install ComfyUI Manager
ComfyUI Manager is the community extension that handles custom nodes, missing-node installs, and one-click updates. Without it, managing video workflows is painful; with it, you can install missing nodes automatically when you load someone else's workflow. Install it through the Manager's own instructions on the ComfyUI-Manager GitHub page, or clone it into ComfyUI/custom_nodes and restart.
Step 3: Download a video model
Model files go into specific folders inside your ComfyUI installation. Video checkpoints go in models/checkpoints, text encoders in models/text_encoders, VAEs in models/vae, and LoRAs in models/loras. Most creators use the Hugging Face CLI for downloads:
huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir models/checkpoints/wan22-5b
Which model should you grab first?
- Wan 2.2 TI2V-5B (best for 8–12 GB cards): Apache 2.0 licensed, accepts text or an image as input, generates about 5 seconds (121 frames at 24 fps). The largest ecosystem of community LoRAs and workflows in open video. No native audio.
- HunyuanVideo 1.5 (best for 14–24 GB cards): 8.3 billion parameters, runs with offloading at around 14 GB, ships first-party distilled variants. Custom license — read the terms on the model card. No audio.
- LTX-2.5 (best if audio matters): a 22B dual-stream diffusion transformer from Lightricks that generates synchronized video *and* audio in a single pass — dialogue, ambience, and effects aligned to the picture. Clips run 2–10 seconds at 720p or 1080p at 24, 25, or 50 fps, with native multi-shot generation that holds character and voice across cuts. Open weights are free for research and for commercial use by companies under 10 million USD in annual revenue — read the LTX Community License on the model card before commercial use.
Update ComfyUI to version 0.32.0 or newer before running LTX-2.5; that version carries the first-party LTX-2.5 workflow templates.
Step 4: Load an official workflow
Do not build your first video graph from scratch. ComfyUI ships official example workflows for supported models, and comfy.org publishes template workflows for LTX-2.5 covering text-to-video, image-to-video, and first-and-last-frame interpolation. Load the template for your model through the Templates panel (Workflows > Browse Templates), then right-click any node to change the checkpoint, prompts, or dimensions.
A standard text-to-video graph has four blocks: a model loader for the checkpoint, a text encoder pair (positive and negative conditioning), the sampler chain (KSampler plus an empty latent sized to your target resolution), and a video decode and save block that writes frames at 24 fps.
Step 5: Generate your first text-to-video clip
With your workflow loaded, set the parameters, then press Queue Prompt:
1. Resolution: start at 720p (1280x720). Higher resolutions can wait until you know the workflow works.
2. Frame count: start small — 49 to 81 frames (about 2–5 seconds depending on fps).
3. Steps: 20–30 sampler steps is the standard range for a test render.
4. CFG scale: 3–7 for most video models; LTX workflows often default near 3.
5. Prompting: write prompts like a shot list — subject, action, camera move, lighting, mood. "Drone shot pushing forward over a coastal town at dawn, warm light, gentle waves" beats a poetic paragraph.
Watch the preview nodes. If the motion looks right, scale up frame count and resolution for your final render. On a consumer GPU, expect minutes per clip, not seconds — plan generation time into your session.
Step 6: Try image-to-video
Image-to-video is the most reliable path to a good-looking clip, because the model starts from a composition you already approve. Generate a still with an image model first, then load it in the workflow's Load Image node and prompt the motion: "the figure walks toward the camera, leaves rustle, clouds drift." Keep motion descriptions simple and physical; abstract instructions ("convey nostalgia") degrade quality.
Wan 2.2's TI2V-5B accepts either a text prompt alone or an image plus prompt in a single checkpoint, which makes it the simplest image-to-video starting point on modest hardware.
Step 7: Upscale with SeedVR2
Since ComfyUI 0.28.0, native SeedVR2 image and video upscaling is built into the core — no third-party node pack required. The standard professional practice in 2026 is generate at 720p, then upscale for delivery: add the upscale stage into the same graph (a Load Video or the generated latent, a latent upscale, then a second sampler pass at low denoise strength around 0.3–0.4), or export your clip and upscale it as a separate pass. Save the final file as MP4 with a constant frame rate — variable frame rates from ad-hoc frame exports are a common cause of stuttery playback that is actually a container issue, not a generation issue.
Troubleshooting the usual failures
Out of memory on a model that should fit. Enable model offloading and attention slicing, drop to a quantized or distilled variant, or reduce frame count before touching resolution. Community GGUF quants and acceleration LoRAs (Wan's "lightning" LoRAs cut sampler steps dramatically) are the standard fixes.
Missing nodes when loading a downloaded workflow. ComfyUI Manager's "Install Missing Custom Nodes" resolves this automatically. If it cannot find a node, the workflow is too old or references a private pack — find a newer version.
Colors shift or the clip flickers. Color mismatch usually means the wrong VAE for the checkpoint; use the VAE the model card specifies. Temporal flicker responds to higher step counts and tighter, more concrete motion descriptions.
Generations are slow. That is normal for local video. The headline speeds vendors quote (Lightricks' sub-7-second 10-second 720p clip, for example) are measured on datacenter superchips, not desktop GPUs. For no-hardware workflows, Runpod and Vast.ai both offer ComfyUI templates with the model weights pre-downloaded — you pay per GPU-second instead of buying a card.