The expensive problems in these shorts were not always difficult actions. Sometimes the wrong character reference was attached. Sometimes the first frame already contained the passage’s key reaction. Sometimes usable individual segments revealed a different room when assembled.
This article reviews the August production handbook’s methods for this project. Historical settings, prices, and individual successful outputs are not treated as lasting guarantees.
Before generation: check what is already knowable
The first review calls no video generator. It checks action chains, the starting state of first frames, connections between shots, and the references actually attached.
Matching names do not prove matching assets. Inspect the images and compare checksums with approved files. Confirm face, clothing, and setting versions. For sound, verify the speaker and reference responsibility rather than merely finding a file in an audio slot.
Check concrete relationships: who holds the prop, whether that hand has another simultaneous job, who reacts first, where the preceding shot ends, and which direction the next camera faces. Longer instructions make separate passes over camera, position, props, time, and light more useful for finding conflicts.
The handbook’s checklist changed as films added evidence. Use the current items instead of remembering a historical item count.
Low-resolution tests: use enough time to answer the question
The second review tests risk at lower cost. A short passage may answer an identity or text question. Reveal order, a complex camera move, or the final landing may require a complete sequence rather than a mechanical five-second opening sample.
The project’s baseline approach kept the target model and lowered resolution. Switching models also changes the object being tested. State failure conditions: a changed face, a transferred prop, a frozen pause, or a premature response gives the test more value than a general impression.
Testing still consumes credits. Its purpose is to reject a route before committing to a full delivery, without guaranteeing the next sample will succeed.
Final delivery: high resolution is a new result
A low-resolution pass does not accept a fresh generation at delivery settings. Identity, blocking, sound, and timing can fail again.
Watch continuously for action, conversation, atmosphere, and continuity. Inspect frames for faces, text, hands, and contact. Neither replaces the other. Attractive particles or clothing in one frame may move incorrectly; technically intact frames can still produce an exchange with no human connection.
Motivated reverse angles and reaction shots can cut. A continuous interaction requires checking whether splitting it destroys proportion and momentum. No shot structure makes every decision: Hanabi used an umbrella cut, while Before Nine retained a longer interaction.
Give the multiplier an explicit denominator
The later handbook uses this accounting form:
Rework multiplier = (actual video-generation credits − separate experiment credits − credits for output reused in a finished channel) / the one-pass delivery baseline.
Reused output can mean a rejected picture whose audio entered the film. The cost still exists in the spending ledger. Deductions also need separate definitions so one credit is not subtracted twice.
The baseline concerns the footage that needed generation at its delivery settings. Locally assembled titles, credits, and subtitles do not turn the whole displayed runtime into generated footage. Hanabi runs about 101 seconds on the site but contains roughly 87 seconds of generated film. Its early same-resolution record lists 264 generated seconds, approximately 3.03 times the baseline.
Before comparing multipliers, align delivery resolution, iteration resolution, and film type. Moving trials to low resolution reduces the accounting multiplier even if the number of mistakes stays the same. Also record problems caught before generation, caught in low-resolution tests, and missed until final delivery.
Leave a route another session can follow
A final-route record should list adopted assets, input responsibilities, instruction versions, settings, accepted variations, and remaining limits. Retired branches belong outside the active production list.
Keep the owner’s assembled film as a read-only master. Similar summed source durations do not prove a simple concatenation. Compression or audio work should produce a new file with its source recorded. This batch’s web copies were exported from complete films rather than reconstructed from shots.
Record an avoided action as a workaround and an accepted performance as this sample’s result. The next film must test whether the workflow has improved. Related fault categories appear in An AI Video Failure Library.