PRODUCTION NOTES
Seven Choices in Making a Music Video
with a record of what was rejected
Over seven music videos, the production choices settled into a shape. There are seven axes you can specify per song, each with a default.
This article is the list of what can be chosen. The mechanism side — which models, how the GPU is run, what it costs — is in Mass-producing music videos with MiniMax H3 (Japanese).
The options below mix what was adopted, what was tried and rejected, and what is not implemented yet. They are marked as such, and rejected options carry the reason. The reason something was not chosen is often more useful than the reason something was.
01Window — which 60 seconds
A song runs three to four minutes; the video is 60 seconds. Where you cut decides what the piece becomes.
| Option | What it is | Example / status |
|---|---|---|
| Late window default | Bridge → final chorus → outro. Lets the piece end at its emotional peak | All 7 songs |
| Opening window | From the top through the first chorus. Shows the story beginning | Not used |
| Chorus only | Cut only the chorus | Not used |
| Digest | Stitch sections together in the edit | Rejected — loses musical continuity |
Burned-in lyrics have to match what is actually being sung, so this choice largely fixes the skeleton of the piece.
02How text exists
Whether to put lyrics on screen, and if so, how. This is what moves the work the most.
| Option | What it is | Example / status |
|---|---|---|
| As objects in frame | Spilled paint, a crosswalk, a split-flap display, the negative space itself becomes the letters. Changing the role per cut lifts the whole piece | 未定稿 (Mitei-kou) — highest rated |
| As light | Lighting, a chain of penlights, the spread of a spotlight forms the letters | 叫べ (Sakebe) |
| As brushwork | Written large across sky, flag, or road surface | 折れない旗 (Orenai Hata) / 走り出したいのに (Hashiridashitai noni) |
| ASS subtitles | Mincho, film-subtitle style. Positions cycle; long lines shrink automatically | 響け、夜 (Hibike, Yoru) |
| None | Image only | Every song, in the outro |
"As objects in frame" rated highest. 未定稿 (Mitei-kou), which changed the role in every cut, was the high-water mark.
03Intensity — how much motion per cut
Do not apply the same intensity everywhere. Stillness is what makes movement land.
| Option | What it is | Example / status |
|---|---|---|
| Still | Reduce motion description. One step. Only the hair moves | Highly rated — the step forward in 折れない旗 (Orenai Hata) |
| Long, low-change stillness | Keeps moving while holding a quiet bearing | Not implemented |
| Medium default | Hair, cloth and particles move on their own; the camera stays modest | Proven |
| Motion | Camera work + depth of field + parallax, while keeping the subject readable | Tends toward monotony |
| Full abstraction | Deformation past the point where the subject reads | Rejected as a direction |
04Cut connection
| Option | What it is | Example / status |
|---|---|---|
| Hard cut default | Generate each cut independently and place them in sequence | All 7 songs |
| Seamless | Specify the next cut’s image as the endpoint of the previous one and interpolate. The switch disappears | Implemented, not yet used in production |
The seamless approach was tested on 流転 (Ruten). That record is a separate article.
05Lip-sync allocation
The axis that moves cost the most. Lip-sync generation takes roughly three times as long as the alternative.
| Option | What it is | Example / status |
|---|---|---|
| Key moments only default | Top of the chorus, the title line — two or three cuts | Standard |
| Every sung cut | All cuts where someone is singing | About 3× the cost, and tends toward monotony |
| None | Instrumental and scenery only | Cheapest |
06Art style
| Option | What it is | Example / status |
|---|---|---|
| Flat colour default | Large flat fields, hairline linework, two-tone shadow | 走り出したいのに (Hashiridashitai noni) |
| Inherit the existing cover video | Carry over character design and style for consistency across the site | 響け、夜 (Hibike, Yoru) / 叫べ (Sakebe) |
| Pop art collage | Unbleached paper, geometry, brush strokes, restricted palette | 未定稿 (Mitei-kou) |
| Editorial poster | Bold negative space and geometry, the texture of print | 折れない旗 (Orenai Hata) |
07How adventurous the angle is
| Option | What it is | Example / status |
|---|---|---|
| Standard default | A mix of close, wide and overhead | Most cuts |
| Unusual viewpoint | Inside an instrument, a reflection in a water drop, the distortion of a mirror, the gap in a forest of arms | 叫べ (Sakebe) — generated successfully, not used |
| Visualising the concept | Draw the idea in the lyric itself — a voice becoming wings, a shadow starting to move | 未定稿 (Mitei-kou) — highest rated |
08How a request is made
In practice you pick from these tables and say it in one sentence.
"Make the video for 叫べ (Sakebe). Late window, text burned in as light, lip-sync on the top of the chorus and the closing cut, art style inheriting the cover, still before the chorus, then moving hard once it hits."
With nothing specified, a proposal comes back on the defaults and is confirmed before generation. If a storyboard or prompt design already exists, that takes priority. スカーレット (Scarlet), 未定稿 (Mitei-kou) and 折れない旗 (Orenai Hata) were rated highest when made that way. A design that exists up front beats leaving it to automatic generation.
09What each song established
The seven were not the same procedure repeated. Each one added something.
| Song | Length | Cuts | Lip-sync | What it established |
|---|---|---|---|---|
| 響け、夜を越えて (Hibike, Yoru wo Koete) | 59.4s | 23 | 8 | Automatic key-visual generation, and a review screen that can redo them on the spot |
| スカーレット (Scarlet) | 56.0s | 10 | 3 | Taking a full storyboard as given, and drawing text as objects in frame |
| What Remains | 59.7s | 10 | 2 | A fully automatic path that pulls source material from the music site |
| 折れない旗 (Orenai Hata) | 60.2s | 11 | 3 | Mixing hand-made and generated images. Placing the cover art as the first frame |
| 走り出したいのに (Hashiridashitai noni) | 59.3s | 12 | 3 | A compositional workaround: never put text on a moving surface |
| 未定稿 (Mitei-kou) | 57.8s | 11 | 3 | Changing how text exists in every cut. The high-water mark |
| 叫べ (Sakebe) | 59.4s | 12 | 3 | Designing unusual viewpoints, and fixing the hang in image generation at its root |
10Quirks of the video model
Behaviours hit repeatedly while working with MiniMax H3, and how they are avoided.
| Quirk | What to do |
|---|---|
| Text burned into an image degrades over time, and once it is gone the model invents different letters | Put text on surfaces that do not move (sky, walls). Never on a flowing surface such as a floor |
| When a character reference is attached to lip-sync generation, the image drifts toward that composition | Pass only the image for that cut itself |
| Writing "the camera moves forward" can produce a shot where the subject approaches instead | Write it as the world moving rather than the camera |
| A first-frame image gets stretched when its aspect ratio differs from the canvas | Crop automatically when building the job |
| Generation beyond eight seconds does not finish in any practical time | Keep every cut within eight seconds |
The first is the worst. Burned-in text does not merely degrade over time — once it has broken up and vanished, the model invents different letters in its place. That happened on 走り出したいのに (Hashiridashitai noni) with text placed on a moving floor, and produced the compositional rule that text only goes on surfaces that do not move.
11Rules that came out of failures
The defaults above were not known from the start. They were decided by getting things wrong.
Altered approved deliverables without being asked
Pixel-matching the first frame produced a verdict of "stretched vertically", and the correction squashed it horizontally instead. That comparison cannot distinguish a natural redraw from genuine distortion. The ground truth for how something looks is the approval, not a measurement. Now approved material is not touched unless someone raises an issue.
Took a correlation for a specification (twice)
Aggregating 319 songs produced a perfect correlation — "no song before a certain date has a short clip" — and the conclusion was a platform limitation. In fact that was simply the day production started. The follow-up theory, "old songs cannot be registered", was disproved by a song ten days old. Now correlation decides nothing; only a controlled experiment that varies the target counts. The real cause was a platform bug: songs opened from search fail to register.
Blamed a text error on the video rather than the image (three wasted runs)
A video with extra characters was regenerated three times, but the error was already in the key visual. The cause was approving on overall impression alone. Now burned-in text is checked character by character before it goes to approval.
Left a stalled machine running and stopped a working one (both directions)
A 0% utilisation reading was taken as grounds to stop a perfectly healthy machine early. On another day a genuinely hung machine was left for 48 minutes. Some rental machines report 0% while working. Now health is judged only by whether the record advances, and a runtime ceiling is derived from queue size so that neglect has a bounded cost.
Found the root cause of a 25-minute hang in image generation
Image generation stalled constantly; ten images were on track to take eight hours. The cause was a single line in the log — "waiting for additional input from standard input". Run in the background, the input stream never terminated, so it waited forever. Passing empty input made the same work finish in five minutes, and ten images in thirty-five.
Motion direction collapsed into "push in, pull out"
Applying the same motion description to every cut produced a video that only alternated between close and distant. What was needed was not a category of movement but a design where depth, focus and object interference are integrated into one image. Now no generic template is applied: still and moving intensity is decided per cut, and the balance follows the character of the song. This became the "intensity" axis in section 03.
12Related
Mass-producing music videos with MiniMax H3 — The technical side in full — models, pins, commands, cost. (Japanese)
流転 (Ruten) — Amor manet production notes — A day spent testing the "seamless" connection from section 04. (Japanese)
Production log 2026-08-12 — The day the correlation mistake in section 11 happened. (Japanese)
Suno × MiniMax H3 production notes — The first two days. (Japanese)