FramePack: Make Free AI Videos Up to 60 Seconds Long on a 6GB Laptop GPU

Every AI video tool on the web charges you per clip — and caps you at five or ten seconds. Yesterday's guide showed ComfyUI with Wan 2.2, which is free but wants a strong GPU for its best quality. FramePack solves a different problem: it makes long videos on weak hardware. One photo in, up to 60 seconds of video out, on a laptop GPU with just 6 GB of VRAM — free, open-source, no subscription, no credits, no watermark.

What FramePack is

• Free and open-source (Apache 2.0) — the code costs nothing, runs entirely on your machine, and has over 17,300 stars on GitHub. The project is by Lvmin Zhang, the researcher behind ControlNet, with co-authors from Stanford and MIT, and the method was accepted at NeurIPS 2025.
• It generates video progressively. Instead of rendering a whole clip at once, FramePack predicts the video section by section, like writing a sentence word by word. That is the whole trick.
• The memory stays constant. FramePack compresses the frames it has already made into a fixed-size context, so the work per frame never grows — frame 1,800 costs the same as frame 30. That is how a 13-billion-parameter video model fits in 6 GB of VRAM.
• Long is the point. Most free tools give you 5 seconds. FramePack is built for 30–60 seconds at 30 fps (up to 1,800 frames) — the length a real scene actually needs.

What you need — the honest version

• GPU: an Nvidia RTX 30, 40 or 50 series card with at least 6 GB of VRAM. A laptop RTX 3060 qualifies — that is the point. GTX 10 and 20 series cards are untested, and AMD and Apple Silicon are not supported.
• System: Windows or Linux. There is no Mac version.
• Disk space: keep 30 GB or more free. The video models download automatically from Hugging Face on first run and are large files.
• Time: this is the part other guides skip. On an RTX 4090 desktop, FramePack makes about 1.5 seconds of video per frame with its TeaCache speed-up (2.5 without it). On a laptop 3060 or 3070 Ti it is roughly 4–8 times slower. A 60-second clip is 1,800 frames — do the maths before you start, and run your first test at 5 seconds.

Step 1: Install FramePack (Windows, one click)

1. Go to the official repository at github.com/lllyasviel/FramePack — this GitHub page is the only official FramePack website.
2. Download the one-click package (CUDA 12.6 + PyTorch 2.6) linked in the README, unzip it, run update.bat first, then run.bat. The update step matters — skipping it leaves you on an older version with known bugs.
3. The models download automatically from Hugging Face the first time you run it. That first download is over 30 GB, so leave it running.

Step 2: Install FramePack (Linux)

1. Use an independent Python 3.10, then install PyTorch with CUDA 12.6 and the project requirements:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
   pip install -r requirements.txt

2. Start the interface with python demo_gradio.py. It opens in your browser as a local page. Use demo_gradio_f1.py for the F1 variant, which most people end up using — it is simpler and faster.

Step 3: Run the sanity check before your own video

Do this first, exactly as the README describes, so you know whether your setup works. Use the README's test image of a man and the prompt "The man dances energetically, leaping mid-air with fluid arm swings and quick footwork." with the default settings and TeaCache off. If the result looks roughly like a dancing man, your install is good. If you get a much shorter video than expected — a second or two — that is normal: FramePack builds the video section by section, so wait. More sections keep being generated until the length you set is complete.

Step 4: Make your first video

1. Upload a clear still image — a person, an animal, a landscape. The image sets the scene.
2. Write a short, motion-focused prompt. The README's own examples are the pattern: "The girl dances gracefully, with clear movements, full of charm." Subject first, then motion, then feeling. One clear action beats a complicated paragraph.
3. Set the length — start with 5 seconds for your first test, then try longer once the workflow runs cleanly.
4. Watch the latent preview as it builds. Because generation is progressive, you see frames appear while it works, so a bad prompt shows itself early and you can stop it.
5. Save the result from the output panel. Repeat with the same image and a different motion to learn what the model does well.

TeaCache: faster, but not free quality

TeaCache is the speed switch in the interface, and the README is unusually honest about it: it is not lossless, and about 30% of users get noticeably different — sometimes worse — results with it on. The recommended habit is to use TeaCache to test ideas quickly, then turn it off for the final version you actually keep. The same caution applies to the other speed tricks (quantisation, sage-attention): fast for drafts, full quality for finals.

Writing prompts that work

FramePack's author includes a ChatGPT template in the README for exactly this: tell ChatGPT it writes short, motion-focused prompts for animating images, then send it an image and get back one concise prompt describing the motion. Prefer big, clear movements — dancing, jumping, running, spinning — over small, subtle ones. If something in the image can dance, make it dance. And keep prompts short: a single sentence with a clear subject and one motion almost always beats a long, detailed paragraph.

Do not get scammed: there is no FramePack website

The README says this in bold, and it is worth repeating: the GitHub repository is the only official FramePack site, and the authors run no web services at all. Domains like framepack.co, framepack.ai and framepack.pro are fake. Do not pay them, do not download from them. There is also no free in-browser demo — the free route is the local install above, on your own GPU. Anyone selling you "FramePack online" is selling you something else.

What it cannot do

This is a research-grade tool, not a polished product. It is image-to-video only — you always start from a still image, there is no pure text-to-video mode in the main demo and no audio at all. Clips are modest resolution rather than HD, and the underlying HunyuanVideo 13B model dates from late 2024, so human motion and fine detail will not match a paid service like Sora or Veo. You will throw away more takes than you keep, and a 60-second clip on a laptop GPU takes hours, not minutes. What you get in return is the one thing no paid tool offers: unlimited long-form video, on your own machine, for the price of electricity.

Bottom line

If yesterday's ComfyUI guide was the most capable free video setup, FramePack is the most accessible one: 60-second clips on a 6 GB laptop GPU, a one-click Windows installer, and no account, queue or subscription anywhere in the process. The price is patience — slow generation on modest hardware — and honest expectations about quality. If you have a gaming laptop with an RTX 3060 or better, this is the weekend project that turns a single photo into a minute of moving video, free, as many times as you like.

Sources: the FramePack GitHub README, checked live 15 November 2026 (17,356 stars, Apache 2.0), and the FramePack paper (arXiv:2504.12626). Hero photo: on the film set of "Wróżby kumaka" in Gdańsk, by Artur Andrzej, CC BY-SA 3.0, via Wikimedia Commons.