FramePack
Apache-2.0 video generator that builds long clips from one image and a prompt on GPUs from 6GB.
FramePack generates video section by section with constant memory use.
FramePack is desktop software and a model that turns an image and a text prompt into video, generated a section at a time. It suits creators with an NVIDIA RTX card who want long clips without a data-centre GPU.
What it is
It compresses earlier frames to a fixed length, so memory use stays the same however long the video gets. A Windows one-click package and a Gradio demo are provided, and frames appear progressively as they finish.
Who it's for
It suits hobbyists and developers with an RTX 30, 40 or 50 series card. It does not suit older GTX cards or Macs, and the author warns that lookalike sites such as framepack.co are fake.
Key features
Constant memory use
Context compression keeps workload flat as clips get longer.
6GB minimum
The README says 6GB of GPU memory handles a one-minute clip at 30fps.
One-click Windows pack
Download, extract, run update.bat, then run.bat.
Progressive output
Frames appear section by section while generation continues.
TeaCache option
Faster runs, with the README warning it can change results.
Gradio app
Linux users start demo_gradio.py after a pip install.
Up and running in 3 steps
Our honest verdict
What's great
- Needs only 6GB of GPU memory for a one-minute 30fps clip with the 13B model, per the README.
- Windows one-click package avoids hand-building a Python environment.
- Apache-2.0 licence on the repository.
- Shows early sections while later ones render, so you can stop a bad run early.
Where it falls short
- Over 30GB of models download from Hugging Face on first run.
- Requires an NVIDIA RTX 30, 40 or 50 series card with fp16 and bf16 support; older GTX cards are not listed as supported.
- On a 4090 the README quotes 1.5 to 2.5 seconds per frame, and laptop cards run four to eight times slower.
- The README warns that TeaCache and Sage-attention can change output, so final renders should skip them.
Pricing & license
Free under Apache-2.0. The only cost is your own GPU and disk space; the README says the author runs no paid service.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
What GPU does FramePack need?
An NVIDIA RTX 30, 40 or 50 series card with fp16 and bf16 support and at least 6GB of memory, on Linux or Windows.
Is framepack.co the official site?
No. The README says the GitHub repo is the only official site, that the author runs no web service, and that lookalike sites are spam or fake.
How fast is FramePack?
The README quotes 2.5 seconds per frame on an RTX 4090 unoptimised and 1.5 with TeaCache, with laptop cards four to eight times slower.
