"It looks nice" is not an evaluation. A film generated from text can be judged on five dimensions, each with a check that a non-specialist can perform in minutes and each pointing at a different stage of the pipeline when it fails. Scoring all five before regenerating anything saves money, because most failures live in the screenplay and cost nothing to fix there.

1. Story fidelity

List the events of the story in order: the fox leaves home, meets the bear, loses the lantern, finds the river. Does each appear as a scene, in that order? Are the character lines the ones the author wrote, or paraphrases? Does the narration add information not in the text, such as a moral the author never stated? Every invention is a defect, even a pleasant one, because the person asked for their story and not the model's.

Fidelity fails in two typical ways. Compression: the target length was too short and the director cut events but kept sentences, so the film jumps. Invention: the director "improved" the ending. Both are fixed in the screenplay, before any video is made.

2. Character consistency

Take three scenes with the same character and compare face, proportions, costume and palette side by side. If you need a subtitle to recognise them, the film fails this dimension. Consistency failures point at casting: the reference sheet was ambiguous, two characters were merged into one entry, or a scene bypassed the keyframe stage and asked the video model to invent the character. The fix is to regenerate the sheet, not the scenes.

3. Style integrity

The first and last scene should look like the same production: line weight, colour temperature, texture, level of detail. A style guide sentence repeated into every prompt is the mechanism, and its absence shows immediately as a film that opens as watercolour and ends as photorealism. Cinematic video tiers are the usual culprit: they add realistic texture to drawn styles. Check also for borders and text: image models sometimes draw a frame or a caption, and the video model then animates a poster.

4. Sound design

Four checks. Are narrator and characters distinct voices? Is each line spoken by the right voice, matching the character's age and gender? Does music sit under speech rather than over it? Is any line narrated twice, once by the narrator and once by the character? The last is the most common fault and the easiest to fix: it is a screenplay rule, not an audio one.

Listen once with the picture off. Speech problems that are masked by the image become obvious: mispronounced names, a narrator that outruns its scene, a music bed that pumps under every line.

5. Pacing

Scene durations should follow the weight of the story: the climax gets the long scene, transitions get the short ones. A film made of thirty eight-second scenes is a metronome, not a rhythm. Check that the director varied durations within the tier's allowed list, that establishing shots are short, and that the ending is not the same length as a transition. Pacing also includes silence: a film with no gap between narration and line has no room to breathe.

Scoring

Score each dimension 0–2: 2 means no defects found, 1 means defects a viewer would notice but forgive, 0 means the dimension fails. Out of 10, a film below 6 should not be regenerated at the video stage; fix the screenplay first, because fidelity, sound and pacing all live there. A film at 6–8 usually needs one cast sheet or two keyframes redone. Above 8, the remaining faults are taste.

Where each fault is fixed

  • Fidelity, sound structure, pacing: the screenplay. Free to change, seconds to regenerate.
  • Consistency: the cast sheets. Cents per sheet, minutes to regenerate.
  • Style, composition: keyframes. Cents per frame.
  • Motion, drift, artefacts: the clips. Dollars per clip; regenerate only the failing scene.

This ordering is the whole point of a staged pipeline: the expensive stage is the last one, and the rubric tells you how far back to go.

A reviewer's routine

Watch once without stopping. Watch again with the story text open and tick events off. Freeze three frames of the protagonist and compare. Listen once with the picture off. Note scene durations from the timeline. The whole routine takes about three times the film's length and produces a score and a list of stages to revisit. Do it before showing the film to anyone else; a first viewer's reaction is dominated by whichever dimension failed worst.

Frequently asked

Can the rubric be automated? Partly. Speech-to-text against the script checks sound fidelity; face embeddings across frames approximate consistency; scene duration variance measures pacing. Fidelity and style still need a person.

What score should a platform aim for by default? A default film from a clean story should reach 7–8 without edits. Below that, the defaults are wrong, not the person's story.

Does a higher video tier raise the score? Only on style and motion. A cinematic tier cannot fix a screenplay that lost the plot.