FramePack by lllyasviel is a free, open source video generation program built on a technique that compresses past frames so the memory needed stays constant however long the video gets. In practice it makes 60 second clips on a 6 GB GPU, showing you the video as it grows. It ships as a small Gradio app with a one-click Windows package and uses a Hunyuan Video based model, with an image as the starting point.

FramePack is for anyone with a modest NVIDIA card who wants to animate a picture into a long clip without a workstation. It is simpler and more focused than Wan2GP, which supports many more models.

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

What you need

Windows or Linux with an NVIDIA RTX 30, 40 or 50 series GPU with at least 6 GB VRAM (fp16 and bf16 support required). 32 GB of system RAM recommended. About 30 GB of disk space for the models, which are downloaded on first run.

How to get and install FramePack

Windows: download the one-click package linked on the GitHub page, extract it and run update.bat once, then run.bat.
Linux or manual: install Python 3.10, clone the repository, install PyTorch with CUDA and run pip install -r requirements.txt, then python demo_gradio.py.
Pinokio also offers FramePack as a one-click install.
On first start the models (over 30 GB) download from Hugging Face; the interface opens in the browser when they are ready.

Download only from the official links on this page. Installers found elsewhere may be out of date or tampered with. ModelTrends.ai hosts no files.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.

Try PirateDiffusion

How to use FramePack

Upload a start image: the first frame of your video. Portraits and clear subjects work best.
Write a short prompt describing the motion, for example "the woman turns and smiles, wind in her hair".
Set the total video length in seconds and the steps (25 is the default). Leave the other settings until you know how your card behaves.
Press Start Generation. The video appears and extends in sections while it renders; you can stop early and keep what is done.
Finished clips are saved in the outputs folder as MP4. Try the TeaCache option for faster renders once you have a result you like.

Tips and common problems

The GPU Memory Preservation slider trades speed for safety; raise it if you get out of memory errors.
Short, concrete motion prompts work better than long descriptive scene prompts; the image already defines the scene.
The F1 variant (demo_gradio_f1.py) generates forward in time and gives more dynamic motion.
Expect a few minutes per second of video on a laptop GPU; keep test renders short.

Models you can run in FramePack

FramePack uses its own Hunyuan Video based model, which it downloads itself. Hunyuan LoRAs from this catalog can be loaded by community forks such as FramePack Studio.

Hunyuan Models

No install required - run AI models on the cloud

No GPU, no downloads, no Python. These services run the same open source models on their own servers, so you can start generating from a browser, Discord or Telegram in a minute.

Try BitVector

Try PirateDiffusion

Links

These links go to the makers' own sites. ModelTrends.ai is not affiliated with them and does not host any software.

Other local apps

How to install ComfyUI

The node-based powerhouse for Stable Diffusion, Flux, Wan video and almost every new open source model. Free, open source, for Windows, Mac and Linux.