The generation rhythm

First one sample shot per scene, reviewed all together; then the rest of the film in production batches, where each remaining shot gets one new take in each batch.

The sample batch (decided): one sample shot per scene, the one that tells most about it. That is its hero shot, or its riskiest: a readable face, a speaking line, hands on the product, product physics, or a method new to the project. Where a scene is one continuous generation, that generation is its sample. Where a scene holds the riskiest generation of a method family, or a face's or place's first production shot, that is its sample, so the review sees every risk early. Your shot plan names each sample and why.

• Two 480p drafts per sample shot (decided): review both, take the better one forward, and log how often the second wins. For one key reaction per film, the two drafts may be two different readings.

• Compare each sample with its scene's reference for the bar before the review.

Inside a batch, shots run side by side. Submit them together, the hero shot first; no shot waits for another's review. A check that could change other jobs of its kind (whether a 1080p finish keeps its draft's motion, whether draft mode accepts an audio reference) runs first and alone.

Your first pass, take by take: review every take as it lands, against the treatment, the approved anchors and the plan's intent, with the full-size checks (see Video with Seedance → Take selection and review). Name every failure with its level and its one next change. Measure the camera dose on the sample batch (see Camera and lens → Measuring the dose on the sample batch).

The batch review. Then judge the whole batch together, in one sitting, from the report below. The takes sit in the round's Videos folder (failed takes in Archive) under their file names, in plan order. Give all your notes at once. The first pass names each take's faults; the batch review sees the takes side by side, so no note is given shot by shot. While the review happens, free work goes on (cut lists, prompts for shots the notes can't touch); nothing is spent. Then apply the notes: restate each note's scope in one line and change only what it named (see The stages and gates → Handling notes).

Production batches (decided): then the remaining shots, each getting one new take in each batch. A take is a 480p draft, given its 1080p finish in the same batch only after it passes your review and the tool's report (see Planning, budgets and takes → Draft small, then analyse), or one direct 1080p take where the plan says so. A finish is not a new take. Each later batch holds the shots without a keeper, each new take carrying one diagnosed change, plus any shot the notes added. Repeat until every shot has a keeper or reaches a stop rule.

• Never spend credits on a take with the same unresolved problem.

• A shot that keeps failing goes into the next batch review with its diagnosis and options: a simpler plan, a split into real cuts, another method. It keeps failing when the same failure comes twice, when it passes its expected figure without a keeper, or when each fix moves the failure, so the plan is the cause. The other shots go on.

• Near a deadline (decided): keep making batches until the deadline, each one reviewed. A deadline never switches a shot to a simpler method by itself; that is your choice.

Stills follow the same rhythm where it fits: one candidate per asset first, each passing your own full-size review, then one batch review. In the face motion tests and the place motion checks, each candidate gets one take in each batch. Voice lines take at least two generations each; Suno at most two Creates per direction.

The order of making:

1. the treatment approved;

2. voices cast and lines recorded, alongside the plan;

3. the plan and the anchors budget approved;

4. the anchors built, and the place motion checks;

5. one anchors presentation, approved by the client;

6. internal frames, only where a method names one;

7. video: the sample batch, its review, the production batches;

8. the rough cut;

9. music options against the cut.

No video for the film runs before the anchors approval, except the face motion tests and the place motion checks.

The dated schedule: work back from the delivery date. Put in each client checkpoint with its send and reply dates, the anchors, the sample batch and each production batch (generation days and review days), finishing and delivery, with a buffer of about 20% of the time left [best available, not yet validated]. If the deadline can't hold the batches the film needs, say so in the plan, with what would have to give (fewer batches, a shorter film, a later date). Never skip a batch review to save time.

{Sample batch | Production batch n}, {date}: {n} shots, {k} keepers.
Scene {n}, {shot}, {what it shows}: keeper. I saw {…}. In {s}, out {s}.
Scene {n}, {shot}: not yet. I saw {…}. Diagnosis: {level and cause}. Next take: {the one change}.
Keeps failing: {shot}: {take by take, what I saw and changed}. Diagnosis: {…}. Options: {…}. Recommended: {…}.
Before each finish: {tool}, quoted as the tool's: {intent lines met / partly / not met}.
Running assembly: {cut}: light {holds | jumps}, colour {…}, wardrobe {…}; average shot length {s} against {range}.
Spent this batch: {cr}. Round so far: {cr} of {expected} / {ceiling}. Next batch: {shots}, about {cr}.

The running assembly is in Editing → The running assembly after every batch.

Sources

• Higgsfield's katana workflow (the riskiest job first); Higgsfield's own team, which runs several takes of a prompt at once and builds scenes from the best moments (the reason for two drafts on each sample).