FastFrames-S: the 1.3B model most people can run
About 2.6 GB of weights in FP16. It downloads in minutes, fits beside the encoder on a consumer card, and is the right place to iterate before you spend a GPU-hour on the big one.
FastFrames-S is the 1.3B checkpoint, and for most people it is the one that actually gets used. This page is about what it can do, what it cannot, and how to get a first clip out of it tonight.
The numbers that matter
Thirteen hundred million parameters in FP16 are about 2.6 GB of weights. That is small enough to sit beside the text encoder and the decoder on a consumer card, and small enough that you can keep the model resident in memory while you change a prompt and run again.
The second number that matters is download time: a couple of gigabytes is minutes, not an afternoon. That changes how you work. With the large model every generation is a decision; with this one it is a keystroke.
What you give up
Small video models agree faster. Expect fewer moving parts that hold together across a long clip, less fine texture, and more sensitivity to prompts that ask for many things at once. The practical response is not to fight it: write prompts with one subject, one action and one camera move, and keep clips short. The prompt builder exists for exactly this.
You also give up the ability to rescue a vague idea by brute force. If a prompt is confused, the large model may resolve it into something beautiful; the small model will usually resolve it into something literal.
A good first hour
- Confirm your card really has 12 GB or more free and a recent CUDA driver — check the model card for the supported stack rather than trusting a forum post.
- Download
apple/FastFrames-Sand keep the cache on a local disk. - Run the inference snippet the model card provides, unmodified, on the default resolution and a short clip. Change one thing at a time from there.
- If you hit an out-of-memory error, cut resolution before cutting anything else — memory scales with the square of resolution.
- Once one clip works end to end, try a full local setup for repeatable runs.
The order matters. People lose evenings by tuning prompts while the environment is still broken, then blame the model.
When to move up to 14B
Move up when the small model's failures are about quality rather than about ideas. If you can already see that the composition is right and only the detail or the motion is weak, the larger checkpoint will help. If the composition is wrong, the larger checkpoint will produce a more expensive wrong composition.
The FastFrames-L guide covers the VRAM situation in detail: roughly 28 GB of weights at FP16, and what it takes to squeeze that onto a 24 GB card.
One thing this page will not tell you
It will not give you a benchmark score. There are no vendor-neutral speed numbers for this release that we have verified, and inventing one would be worse than leaving the gap. What the page does instead is give you the arithmetic and the order of the levers, which is what you need to predict your own machine's behaviour.
Frequently asked
How much VRAM does FastFrames-S need?
The 1.3B weights are about 2.6 GB in FP16. With the text encoder, decoder and activations, a card with 12 to 16 GB is a comfortable starting point; clip length and resolution decide the rest.
Does FastFrames-S give worse results?
It gives simpler results: less fine detail, and less coherence over longer clips. For prompt iteration, storyboards and short social clips that trade is usually worth it, and the large model is a click away when you need the final pass.
How long does FastFrames-S take to download?
A couple of gigabytes of weights, so minutes on a normal connection rather than an afternoon. Keep at least 20 GB of free disk for the weights and the cache copies.
Can I run FastFrames-S without a GPU?
Practically, no: video diffusion on CPU is measured in hours per clip. Use a community Space for a quick look, then a GPU for real work. The where-to-try page lists the demos that are live.