Cloud video tools bill you per second, add watermarks on cheaper plans and can change their terms overnight. Open-weight video models now run on a single consumer graphics card, so you can make clips on your own machine for the cost of electricity. This guide walks through the whole process, from checking your hardware to rendering your first text-to-video and image-to-video clips, using free software only.
| Quick answer: to generate AI videos locally, install the free ComfyUI app, download an open video model such as Wan 2.2 5B (fits about 8GB of VRAM with offloading) or LTX-2.3 (16GB recommended), put the model files in ComfyUI’s model folders, open the matching video template, type a prompt and press Run. For image-to-video, load a starting image into the template’s image node before you run it. |
Table of Contents
What you need before you start
Three things decide whether this works: your graphics card’s video memory (VRAM), your system RAM and free disk space. Video models are far hungrier than chat models, and they lean on system memory when VRAM runs out, so a balanced machine matters more than a fast processor.
| Component | Minimum that works | Comfortable | Why it matters |
|---|---|---|---|
| GPU VRAM | 8GB (Wan 2.2 5B with offloading) | 16GB or more | Decides which models load at all |
| System RAM | 32GB | 64GB | Models offload to RAM when VRAM is full |
| Disk space | 50GB free | 200GB+ on an SSD | Each model and text encoder is several GB |
| Graphics driver | Current NVIDIA driver | Current NVIDIA driver | Most workflows are built and tested on NVIDIA cards |
The 8GB figure comes from ComfyUI’s own Wan 2.2 documentation, which says the 5B model fits 8GB with native offloading. NVIDIA’s RTX video generation guide recommends 16GB of VRAM and 64GB of system RAM for comfortable local video work. If your card is below 8GB, read our best local AI video generator guide, which covers the lightest models, or rent a cloud GPU by the hour.
How to generate AI videos locally, step by step
How we compare: the steps below follow ComfyUI’s official documentation for its built-in video templates. File names and folders are taken from that documentation, so they match what the app expects. We use Wan 2.2 5B as the worked example because it has the lowest stated memory need and an Apache 2.0 licence.
Step 1: Install ComfyUI
Download the Desktop version from ComfyUI’s download page. It is a one-click installer with automatic updates, and ComfyUI is free and open source under GPL-3.0. A Portable build and a manual install from GitHub are also available if you prefer to manage Python yourself. Put ComfyUI on a drive with plenty of free space, because the model files are large. Use the Desktop app rather than Comfy Cloud: the cloud service is a separate paid product, and only the Desktop, Portable and manual versions run on your own hardware.
Before downloading a video model, run one of the built-in image templates. NVIDIA’s ComfyUI guide suggests starting with a text-to-image template, and a successful image tells you the app, driver and graphics card are working before you spend time on multi-gigabyte video files.
Step 2: Pick a model that fits your card
Match the model to your VRAM before you download anything. Starting with a model that does not fit is the most common reason a first attempt fails.
| Model | Practical VRAM | Licence | Good first choice if |
|---|---|---|---|
| Wan 2.2 TI2V-5B | About 8GB with offloading | Apache 2.0 | You have an 8-12GB card or need clear commercial terms |
| LTX-2.3 | 16GB (FP8) | LTX-2 Community | You want the best quality plus generated audio |
| HunyuanVideo 1.5 | 14GB with offloading | Tencent Hunyuan Community | You want realistic people on a mid-range card |
| Wan 2.2 A14B (T2V or I2V) | 24GB and up, or quantized builds | Apache 2.0 | You have a high-end card and want peak open quality |
Step 3: Download the model files into the right folders
A video workflow needs three kinds of file: the video model itself, a text encoder that reads your prompt and a VAE that turns the result into frames. For Wan 2.2 5B, ComfyUI’s Wan 2.2 guide lists wan2.2_ti2v_5B_fp16.safetensors for models/diffusion_models, umt5_xxl_fp8_e4m3fn_scaled.safetensors for models/text_encoders and wan2.2_vae.safetensors for models/vae. Restart ComfyUI after adding files so it can see them.
Prefer LTX-2.3? If you have 16GB of VRAM, LTX-2.3 is the higher-quality route and the only open model that generates matching audio with the video. Open its template from the same browser; when a template needs files you do not have, ComfyUI lists the missing models so you can fetch them, and the FP8 and distilled versions are the ones to choose on 16GB cards. The steps that follow are the same for either model.
Step 4: Open the template and write a prompt
Open the template browser and choose the Wan 2.2 5B video template. Check that the Load Diffusion Model, Load CLIP and Load VAE nodes show the files you downloaded. Then edit the positive prompt in the text encoder node. Describe the subject, the action, the camera and the lighting in plain sentences, for example: “a red kite rising over a windy beach at sunset, handheld camera, slow pan left”. Use the negative prompt for what you do not want, such as blur or distorted hands.
Step 5: Set length and size, then run
The latent video node controls the resolution and the number of frames, set by its length value. Start small: a short clip at a modest resolution renders fast and tells you whether the prompt works. Click Run or press Ctrl+Enter (Cmd+Enter on a Mac). The finished clip appears in the output node and is saved to ComfyUI’s output folder.
Step 6: Turn an image into a video
The same 5B template handles image-to-video. ComfyUI’s guide says to enable the Load Image node with Ctrl+B, upload your starting picture, describe the motion you want in the prompt and run it again. The 14B image-to-video model uses its own template with two model files, a high-noise and a low-noise version, and needs much more memory. For more ideas on turning stills into clips, see our guide to free AI image-to-video generators.
How to fix common problems
Out-of-memory errors: lower the resolution or the frame count first, then switch to an FP8 or quantized GGUF build of the model. Close other apps that use the GPU, including browsers with hardware acceleration. Very slow renders: check that ComfyUI is using your graphics card rather than the processor, and try a distilled model version, which needs fewer steps. Missing nodes or red boxes: update ComfyUI, because new models usually need the latest version, and use the Manager to install any custom nodes a shared workflow asks for.
Choppy or short clips: generate at a lower frame rate and use a frame-interpolation step afterwards, then upscale the result. Both are separate workflows in ComfyUI. Results that ignore the prompt: shorten it, put the main subject first and describe one action rather than several.
Check the licence before you publish
Running a model locally does not automatically mean you may sell what it makes. Wan 2.2 is Apache 2.0, which allows commercial use. LTX-2.3’s community licence allows free commercial use only for organizations under a revenue threshold, and the Tencent Hunyuan licence excludes the European Union, the United Kingdom and South Korea. Our open source AI video generator guide compares the terms in more detail.
How one wedding videographer made her first local clips
Illustrative example. Rafaela Mendes films weddings near Porto and wanted short animated openers for her highlight reels without uploading clients’ photos to a cloud service. Her editing PC already had a 12GB graphics card and 64GB of RAM.
She installed ComfyUI Desktop, started with Wan 2.2 5B because it fit her card and its Apache 2.0 licence was clear, and used the image-to-video template on a few venue photos. Her first renders failed with memory errors at a high resolution, so she dropped the size and frame count, then upscaled the best clips afterwards. Within a weekend she had a repeatable workflow, and the couples’ photos never left her machine.
If you want the bigger picture of running models on your own hardware, start with our local AI guide, or see how to run AI models offline for a setup with no internet connection at all.
Frequently Asked Questions
Can I generate AI videos locally for free?
Yes. ComfyUI is free and open source, and models such as Wan 2.2, LTX-2.3 and HunyuanVideo 1.5 are free downloads. Your only costs are the hardware you already own and electricity. Check each model’s licence before using the output commercially.
How much VRAM do I need to generate AI video locally?
About 8GB is the practical minimum: ComfyUI says Wan 2.2 5B fits 8GB with offloading. 16GB is far more comfortable and is what NVIDIA’s video generation guide recommends, together with 64GB of system RAM. Larger 14B models need 24GB or more unless you use quantized builds.
Can I do image-to-video locally?
Yes. The Wan 2.2 5B template in ComfyUI handles both text-to-video and image-to-video: enable the Load Image node, upload a starting picture and describe the motion in your prompt. Larger image-to-video models give better results but need much more memory.
How long does it take to generate an AI video locally?
It depends on the model, resolution, clip length and graphics card. A short low-resolution test can take a few minutes on a mid-range card, while long high-resolution clips can take much longer. Start small, then raise the size once the prompt works.
Can I generate AI videos locally on a Mac?
ComfyUI can run on Apple silicon Macs, but most video workflows are built and tested on NVIDIA graphics cards and run more slowly on a Mac. A Mac with plenty of unified memory can load larger models, but expect longer render times than on a comparable NVIDIA card.
Do I need an internet connection to generate video locally?
Only to download ComfyUI and the model files. Once they are on your drive, generation runs entirely on your own computer, and prompts and images never leave it. Our guide to running AI models offline covers a fully disconnected setup.
Conclusion
To generate AI videos locally, you need a capable graphics card, the free ComfyUI app and an open model that fits your memory. Start with Wan 2.2 5B on 8-12GB cards or LTX-2.3 on 16GB, keep your first clips short, and raise quality once the workflow is reliable.
Check each model’s licence before you publish, and keep your source images and footage on your own machine. The fastest way to learn is one short clip, one change at a time.
