Judging motion honestly

A take is judged against what the plan meant it to do, never against a storyboard frame. "No cut, no morph, no dissolve" is not a pass.

Against the intent [proven]: the treatment, the approved anchors and the plan's intent for the shot: the camera's path and landing, the acting task, the light events, the start and end states.

Three decided rules set what the intent must include:

• The motion standard (decided; its home is Video with Seedance → Take selection and review): every shot stunning and dynamic next to the treatment, the anchors and the plan, with camera, acting, body, hair, hands, face, the product's texture and light moving together; each shot distinct in acting and camera, yet fitting the rest. A slow or static camera where the plan wants motion fails, whatever else holds. A named lock is judged as a lock, with the talent and the place still alive.

• The place stays alive (decided): 2–4 natural background motions outdoors or wherever the place has them, 0–1 in a quiet room, and every take is checked for a frozen background (see Making it feel real → The place stays alive).

• The camera during speech (decided): in a speaking shot, the camera breathes or pushes very slowly during the words, with bigger moves before or after the line (see Speaking shots and lip sync → The camera during speech).

One kind of evidence per pass:

• Frames every 0.5 s, plus the first and last, for counts, screen sides, eyelines, reaction timing and cut points (steps in Do it yourself below). Write notes as "I saw , so I would ".

• Measured [best available, not yet validated]: the camera's travel, stalls, jumps and drift from a tracker, reported beside the verdict and never as a pass mark (see Camera and lens → Measuring the dose on the sample batch).

• Reported by a tool: a video-understanding tool's read of what moves and how it feels.

• Your eye: you watch every take that reaches the cut, at speed, and decide.

What a video-understanding tool can tell you [best available, not yet validated]. Use Gemini first (Gemini 3.8 Flash, or 3.1 Pro for a hard read such as a dense acting scene), which hears sound and answers questions, and Higgsfield's video analysis as the fallback, which takes no question and returns a fixed scene-by-scene read timed in whole seconds. Keep each clip for Higgsfield's analysis under about 30 s: its accuracy falls on longer clips.

How to ask:

• One take per request, cut to the window that will play, so the report talks only about what will be seen. Sample at 5 frames a second for a take, up to 10 for a short action or hand moment. That rate is for the tool; your own frame export is every 0.5 s, for checking by eye. How each draft's window is cut and reported before its 1080p finish, with the export command and the costs: Planning, budgets and takes → Draft small, then analyse.

• Write the plan's intent into the question, one line each:

  • the camera move, with its trigger, its size and the visible change of framing;

  • the acting task and each planned reaction;

  • the light events;

  • the named background motions;

  • the start and end states, and the planned cut;

  • what must not happen.

• Then ask:

  • what actually moves, and when;

  • whether it moves like a real rig or floats;

  • whether anything changes shape or count;

  • whether any light has no source;

  • whether the background freezes or slides as one flat card;

  • what is heard;

  • each intent line as met, partly met or not met, with times.

• Higgsfield's read can't take these questions: map its scenes onto them yourself and mark the rest "not reported".

• Check every claim that would change a decision on the frames at the time the tool gives, within a second either side. If the frames disagree, the frames win.

• Treat counts, text and names in a report as unverified. Models invent confident detail on long or busy clips.

• Upload client material only where the client's non-disclosure agreement (NDA) allows, and to Google only through a paid API key, the key that opens Google's programming interface on a paid plan (on the free tier, Google's terms let it use what you upload to improve its products).

Do it yourself

Export frames every 0.5 s, plus the last frame (ffmpeg)

1. Export a frame every half second; the first file is the take's first frame:

ffmpeg -i take.mp4 -vf fps=2 f_%04d.png

1. Export the last frame:

ffmpeg -i take.mp4 -update 1 last.png

Ask Gemini about a take (the Gemini API, with a paid key)

If you don't write code, whoever set up your Gemini key sends the request: hand them the window, the frame rate and your question file.

1. Cut the window that will play from the draft, as an H.264 MP4 (the command is in Planning, budgets and takes → Draft small, then analyse → Do it yourself). Send one take per request, so the report talks only about what will play.

2. Choose the model and the key. Use Gemini 3.8 Flash; use Gemini 3.1 Pro for a hard read (a dense acting scene, or when Flash and your frames disagree). Run it on a paid Google API key: on the free tier, Google's terms let it use what you upload. Upload client material only where the client's NDA allows.

3. Upload the window through the Gemini API's file upload, with your paid key.

4. Set the video's frame rate to 5 frames a second, up to 10 for a short action or hand moment. Gemini's default of 1 frame a second is enough only for a slow reference. Set the media resolution to high and the temperature low (0.2), so the answer stays factual.

5. Write the question in a plain text file: paste the text below, fill it from your shot plan's row for the shot, and delete the lines that don't apply. It holds the plan's intent lines and the questions under "How to ask" above. Arabic lines go inside «…», exactly as approved.

You are reviewing one generated take for a film director. Judge it only against the intent below, and describe only what you can see or hear.
Shot {shot}, {seconds} s, {aspect}. The clip holds only the stretch of the take that will play in the film.
INTENT FROM THE PLAN:
- Camera: {the move, what sets it off, how much the framing changes and where it lands | a locked frame: the camera does not move or zoom}.
- Acting: {what the person does and why}; {each planned reaction, its type and its time}.
- Light: {the source, and any light event | no light change}.
- Background: {the live motions the plan names | a quiet room: only {one motion}}.
- Start: {…}. End: {…}. The edit {cuts while it moves | cuts after a hold}.
- Must not happen: {…}.
Answer with times in seconds, to one decimal:
1. CAMERA: what moves, when it starts and what starts it, any pause or stop, how much the framing changes, where it lands, how the last second moves. Does it move like a real rig with weight, or float at one even speed? On a locked shot, any drift or zoom?
2. ACTING, only where a face reads: what the person does, beat by beat; anything overplayed, rushed or frozen; does each reaction land once, at its planned time?
3. BODY AND HAIR: do hair and hands move naturally; do moves gather, travel and settle; does any limb bend or merge?
4. OBJECTS: does anything change shape, size, colour, count or place, or appear or vanish?
5. LIGHT: any flare, glow, flash, self-lit surface, or light with no source?
6. BACKGROUND: does the place stay the same; does each named motion keep going to the end; do the near and far layers part as the camera travels, or slide together as one flat card?
7. SOUND: what is heard; any music, voices, swell, whoosh or fade?
8. VERDICT: each intent line as met, partly or not met, with the time of any problem.
Write "unclear" where you cannot tell. Do not judge faces, labels or text: they are checked separately at full size.

1. Read the answer from item 8 first: every intent line marked met, partly or not met. Check every claim that would change a decision on your frames, as "How to ask" says above; if the frames disagree, the frames win.

2. Save the report in the take's entry in your take log: copy the lines that matter, quoted as Gemini's, with the full answer below them.

Sources

• Google's Gemini video-understanding guide.

• Higgsfield's video analysis.