Browser tool

FastFrames prompt builder

Five slots, one action: subject, action, camera, light, look. Fill them in, get a prompt that a text-to-video model can actually hold together across a clip.

Fill in a subject and an action, then press Build prompt.

The wording is assembled in your browser — nothing is sent anywhere, and no video is generated here. Paste the result into a FastFrames demo or your own pipeline.

Text-to-video prompts fail in a different way from text-to-image prompts. An image is a single frame, so describing five things is fine — the model simply puts them all in one picture. A clip has to hold those five things together across dozens of frames, and the more you ask for, the faster it falls apart. The structure below is the one that survives that.

The five slots

A FastFrames prompt that works has exactly five pieces of information:

  1. Subject — one thing, described concretely. “A brass desk lamp with a dented shade” beats “a nice lamp” because the model has something specific to render.
  2. One action — a single change over time. “Flickering on and settling” is one action. “Flickering, then someone walks in, then the camera pans to a window” is three, and the third one will be the one that breaks.
  3. Camera — stated explicitly, even when it is a locked-off shot. If you do not say it, the model guesses, and guessing is where motion artefacts come from.
  4. Light — one source or one condition. Light is what makes a generated clip look expensive rather than plastic.
  5. Look — film stock, rendering style or lens character. One choice, not a list of adjectives.

The builder above assembles exactly this, in this order. Nothing about it is magic: the value is that it refuses to let you write a prompt with four actions in it.

Three prompts that follow the rule

a brass desk lamp with a dented shade, flickering on and settling into a steady glow,
slow push in, single hard key light with deep shadows, photorealistic 35mm film grain
a ceramic teapot with a chipped lid, slowly rotating on a wooden table,
static locked-off shot, soft window light from the left, clean product photography
a paper boat on a wet pavement, drifting slowly into a puddle,
low angle looking up, flat overcast daylight, documentary handheld footage

Notice what is absent: no “cinematic”, no “8K”, no “masterpiece”. Those words describe quality you want rather than content you want, and they compete with the words that actually decide the frame.

The four mistakes

Two subjects. Two characters means the model has to keep two things consistent, and small checkpoints in particular will merge them. Split it into two clips.

An action that needs a second action. “Opens the door and turns around” is two beats. Pick the one you care about.

Negation without a negative prompt. “No text, no watermark” inside the positive prompt often produces text and watermarks. Keep the positive prompt about what is there.

A long clip. Motion coherence decays with length. Four seconds that hold together is worth more than ten seconds that dissolve, and it is cheaper on every axis.

Iterate on one word at a time

When a clip disappoints, the instinct is to rewrite the whole prompt. Resist it. Change the camera and generate again, then change only the light. You are looking for the variable that moved the output, and that is only visible when one variable moves. This is also why the 1.3B model matters: at 2.6 GB of weights you can afford twenty attempts, and twenty attempts teach you more about the model than one perfect render.

No GPU yet? Start with a community demo — the where to try page tracks what is live.

Frequently asked

How long should a FastFrames prompt be?

One sentence, usually 15 to 30 words. A subject, one action, a camera behaviour, a light and a look. Longer prompts rarely help because the model can only keep a few constraints consistent across a clip.

Does the prompt builder generate video?

No. It assembles prompt text in your browser and nothing is uploaded. To generate, paste the result into a community demo or your own FastFrames setup.

Should I add words like cinematic or 8K?

Usually not. They describe the quality you want rather than the content you want, and they compete with the words that actually decide the frame. A specific camera move and a specific light do more than a quality adjective.

Do I need a negative prompt?

If your pipeline accepts one, put artefacts there rather than in the positive prompt. Writing “no watermark” inside the positive prompt often has the opposite effect, because you have just named the thing you do not want.