Chapter 19 · Veo 3.1 and MiniMax H3

Yasser Ashour

2 hours ago

10 min read

Part 3 · Motion

Two video models that generate their own sound and cannot switch it off. Veo 3.1, Google's model, suits an action whose sound carries the idea. MiniMax H3 renders at 2K from a keyframe or a mix of references. Neither is a route for an Arabic line. This chapter teaches both in short form and says what to check before a client shot depends on either.

In this chapter

What each model is, its current version, and its role
How to write a clip and its sound for each; templates and two examples
Settings, checks, failures, limits, prices and rights

Before you start. Chapter 13 chooses the route; Chapter 14 teaches the seven elements of a clip, which both models use unchanged.

19.1 What it is

Veo 3.1

Veo 3.1 is Google's current Veo (checked 30 Sep 2026): 4, 6 or 8 seconds at 24 frames a second, 16:9 or 9:16, with generated sound that is always on. Google lists it as a preview on the Gemini API; Google has deprecated its earlier Veo models. It comes in a fast variant (the default), a slower preview variant, and Veo 3.1 Lite. A newer Google model now sits beside it: Gemini Omni Flash 1.1, Google's default video model (19.10; entry in Chapter 20). This chapter teaches Veo 3.1, the Veo the connection offers, whose contract has been read.

Its role (check it on a scratch clip before a client relies on it: 19.7). A sound-led English action, a visible impact whose sound carries the idea, such as a cap twisted shut or a can opened; or a premium wordless plate whose track is stripped afterwards. Google says of speech: "English (EN) is fully supported, but other languages have not been evaluated".

MiniMax H3

H3 (Hailuo 3.0) is MiniMax's video model, announced 31 Jul 2026. MiniMax's documentation offers two video models, H3 and H3 Max, and no newer one (checked 30 Sep 2026). H3 takes text, image, video and audio together; H3 Max is the fast variant, post-trained by fal.ai on H3, at 480P and 768P.

Its role (check it the same way: 19.7). A 2K picture from a keyframe, a plate for delivery at 1080p or better made in one job instead of a plate plus an upscaler; and mixed references (images, footage, audio) in one pass. It is not a speech route, and not for a plate that must arrive silent.

Both are reached through Higgsfield (Chapter 33), which exposes less than the makers' own sites (13.6 gives the map). Name the model in every request: the host's video tool names H3 its default for 2K keyframes or mixed references, so an assistant may pick it unasked, sound included.

19.2 How it reads what you give it

Veo 3.1. Text and a start frame. The frame owns everything that exists; the prompt owns what changes. The connection takes a start frame only (Lite takes a start and an end frame). Google's own site also takes up to three reference images, first-and-last-frame interpolation and extension of Veo clips. Sound cues in the prompt are read: dialogue in quotation marks, effects and ambience described. Google limits the text to 1,024 tokens.

MiniMax H3. Text is required, up to 7,000 characters. Around it: a start frame, an end frame, or both; or references, up to nine images, three videos and three audio clips, twelve files in all. On MiniMax's own API, frames and references cannot mix in one request; whether they can on the connection is not confirmed. Label each input by number and job ("Image 1: the car; reference video 1 gives the motion only"). An audio reference gives voice, its timbre and rhythm; audio cannot be sent alone, and no sentence says a supplied take drives the lips.

19.3 The prompt, in this tool's dialect

Both use the seven elements of Chapter 14: the camera with its rig physics; the subject's action; the world's life; what is protected, by name; the endpoint as a cuttable composition with its hold; the speed profile; earned exclusions (three or fewer; five is the ceiling). Nothing the start frame carries is described again, and the look is never restyled: Google's guide lists style words, which on a clip from a frame only fight it.

Sound is part of every prompt, because the track cannot be switched off. Close the prompt with a sound line that names each sound at its moment and the ambience beneath it. Google's guide has three registers: dialogue in quotation marks, effects described explicitly ("tires screeching loudly"), and ambient noise described. "No music, no voice" is a request, not a switch: check the file.

Sound-led action (Veo). Write the impact and its sound as one physical event with a moment: the click at the lift, the pop at the opening. Keep the clip at 4 seconds.

Speech. An English line goes in quotation marks with who says it, and only where the project has chosen a new performance (Chapter 22). Egyptian and Saudi dialogue is never generated here; it comes from an actor or ElevenLabs.

H3. MiniMax puts camera brackets such as [pan], [zoom] or [static] directly after the description they apply to, and ends its sample prompts with a "Sound:" line. Duration and aspect are settings, not words.

19.4 Templates

Written for this book · template; not run · finished grade · Veo 3.1 or MiniMax H3, a clip from a start frame

Template — a clip with its sound

[Camera: the rig, what it does over the clip, its physics]. [The subject's action as positions and moments]. [The world's life: what moves because the action happens]. [What is protected, by name]. [The endpoint composition and its hold]. [The speed profile]. [Up to three earned exclusions]. Sound: [each sound at its moment, then the ambience]; no music, no voice[, or an English line in quotation marks and who says it].

Written for this book · template; not run · finished grade · MiniMax H3, start and end frames, or references

Template — a keyframe move, or a mixed-reference clip

Image 1 is the start frame [and Image 2 the end frame]: [what each carries, once]. [Or: Image 1 is the (subject) and controls (one job) only; reference video 1 gives the motion only.] [The camera's travel, in seconds]. [The action and its moments]. [What is protected, by name]. [The endpoint and its hold]. [The speed profile]. [Up to three earned exclusions]. Sound: [each sound at its moment]; no music, no voice.

Written for this book · template; not run · test grade · Veo 3.1 or MiniMax H3, a first look

Template — a scratch clip

[The move]. [The action]. [The endpoint]. Sound: [the sound in words]; no music, no voice.

19.5 Examples at the standard

Written for this book · food and drink · Veo 3.1, fast, basic, 4 s, 16:9, sound on, start frame only: an unopened can, a fingertip on its tab · finished grade; not run

Example 19.1 — a can opened, sound first

The camera is locked off on a tripod and does not move. The fingertip lifts the tab: the ring rises to about forty-five degrees over the first second, the scored opening cracks at about the one-second mark with one short bright pop, and a fine mist of carbonation lifts from the opening and thins over the remaining time. Two condensation drops on the can's shoulder run a finger's width down, and the mint leaf on the ledge trembles once at the pop. The can and its blank surface, the fingertip's place on the tab, the ledge and the light stay exactly as in the start frame. The clip ends on the opened can, the tab raised, the mist thinning, the fingertip still at the tab, held still for the last half second. Real time; the lift eases in, the pop is instant. No liquid overflow, no second can, no camera move, no lettering. Sound: the metal tab's click at the lift, one short bright pop, a fizzing hiss that fades over two seconds, a faint kitchen fan hum beneath; no music, no voice.

Why it works: the impact and its sound are written as one event with a moment, so picture and track can land together.
Copy the sound line: every sound at its moment, the ambience beneath, and a refusal of music and voice, because the track cannot be switched off.
Copy the hold: the clip ends on a composition you can cut from.
Settings: 4 s, not the 8 s default, which doubles the price.

Written for this book · automotive and luxury · MiniMax H3, 5 s, 2K, 16:9, one video, sound on, start and end frames: a dark graphite SUV at dawn, lamps off, then lamps on and closer · finished grade; not run

Example 19.2 — the lamps wake, at 2K

Image 1 is the start frame and Image 2 the end frame; both show the same car on the same forecourt, so nothing they carry is described again. The camera is a level slider push-in, travelling about forty centimetres toward the car over the five seconds, easing in over the first second and out over the last, with no shake. The car does not move. Real time; the lamps rise smoothly, without flicker. The headlamps switch on: the thin light signature lights from the inner corner outward from two seconds, reaches full at three and holds. Palm fronds at the far edge stir in a light breeze, and the wet patches on the forecourt begin to catch the lamp glow after three seconds. The car's shape, colour, panel gaps and wheels, and the villa wall, stay exactly as in the two frames; no badge or lettering appears. The clip ends on Image 2's composition, lamps at full, held still for the last second. No people, no second vehicle, no rain, no lens flare. Sound: dawn birdsong and a faint sea wash, a soft relay click as the lamps light; no music, no voice.

Why it works: it describes only the travel between two relatives: the push, the moment the lamps light, how long they take.
Copy the protection: the car's parts are named, so the lamps are the only change.
Check at 200 %: whether the 2K is real detail, and what is on the track (19.7).

19.6 Settings

Setting	Veo 3.1 (fast, preview; Lite)	MiniMax H3 (H3 Max)
Duration	4, 6 or 8 s; the default is 8, so set 4	4 to 15 s, default 5 (Max: 5 to 15)
Quality / resolution	basic (default), high, ultra; fast (default) or preview; Lite has none	2K only (Max: 480p or 768p, default 768p)
Aspect	16:9 or 9:16 (Lite also auto)	auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
Sound	always on; on Lite a switch, default off	always on (Max: not known)
Inputs	start frame (Lite: start and end frame)	start frame, end frame, image, video and audio references
Batch	one	1 to 4, default 1; each video is billed
Not exposed	references, extension, seed, negative prompt, language, audio input	seed, negative prompt, frame rate, sound switch

Write every setting on every job, and ask for the price of the exact shape first.

19.7 Checks before you sign

Probe the returned file before you look at it: size, frame rate, length and audio stream.

The track, by ear. Does the sound sit on the contact frame? Is the track usable as it comes, or ambience you would strip? Did music or a voice arrive that you did not ask for?
Veo: the length you asked for (4 s), the aspect, and the resolution.
H3, at 200 %: is the 2K real detail, or soft pixels? Do identity and geography hold on the key frames?
The start frame's look. Did the clip restyle it?
A file that must arrive silent has its track stripped afterwards; the sound was paid for either way.

19.8 Failures and fixes

You see	Why	Smallest fix
A voice or music appears	the track cannot be switched off and the sound line left room	name each sound and refuse the rest; strip the track in post; or use a tool with a switch (Chapters 16 to 18)
The sound drifts off the contact	the event was vague	write the event and its sound as one moment; shorten the clip
The clip is 8 s and the price doubled	the default	set 4
Arabic speech, or any language but English	not evaluated by Google; not documented by MiniMax	never; an actor or ElevenLabs (Chapters 21 and 22)
The look changed	the prompt restyled the frame	remove style words; the frame owns the look
Soft detail at 2K	the 2K may not be real detail	judge at 200 %; choose another route for that plate
Four videos billed	batch above 1	batch 1
A reference's identity drifts	one reference did two jobs	one job per reference

Retry ladder, smallest lever first: the settings; the sound line; the failing sentence as positions; a shorter clip; another tool.

19.9 Limits, prices and rights

Veo 3.1 (Google, checked 30 Sep 2026). 24 fps; 720p by default, 1080p and 4K only at 8 s (no 4K on Lite); one video per request; extension adds 7 s up to 20 times, on Veo clips at 720p; videos are kept for 2 days; a request takes 11 seconds to 6 minutes; every video carries an invisible SynthID watermark; in the Middle East and North Africa, people generated from a frame are limited to adults. Google's price per second: Veo 3.1 USD 0.40, Fast 0.10 at 720p, Lite 0.05. Google's terms (effective 23 Mar 2026) say it will not claim ownership of generated content and may generate the same or similar content for others.

Veo 3.1 on Higgsfield, 27 Sep 2026	Credits
fast, basic, 4 s / 8 s	16 / 32
preview, basic, 4 s / 8 s	40 / 80
Lite, 4 s / 8 s	6 / 12

MiniMax H3 (MiniMax, checked 30 Sep 2026). 768P or 2K on the maker's API (2K only on the connection); 24 fps (the maker's page of 11 Sep 2026; probe the file); inputs as in 19.2. The maker's price is USD 0.13 a second at 2K and 0.08 at 768P; H3 Max is 0.05 at 480P and 0.08 at 768P; regenerating a 768P clip at 2K costs 0.05 a second on the maker's API. MiniMax's terms of service could not be read as text on the day: read them before a client job.

H3 on Higgsfield, 27 Sep 2026	Credits
5 s / 10 s / 15 s, 2K	10 / 20 / 30
5 s, 2K, batch of four	40
H3 Max, 5 s, 768p / 480p	12.5 / 7.5

Shapes with video or audio references have not been priced; ask on the day (Chapter 34).

What may ship. An accepted plate, with a track you would keep or a track stripped; no Arabic speech from either model; no unreleased client pack on a service whose terms you have not read.

19.10 Version notes

Gemini Omni Flash 1.1 (generally available 27 Aug 2026) is now Google's default video model; Google keeps Veo 3.1 for scene extension, last-frame control and existing pipelines (Chapter 20). Veo 3.1 is the current Veo; its last update was January 2026 (Lite, March 2026). Re-check when the connection lists Omni, and whether Google keeps Veo for those jobs. H3 is the current MiniMax video model. Re-check H3 Max's sound.

Open questions

What Veo returns at high quality against basic at 4 s, and whether the sound lands on the contact frame as written.
Whether H3's 2K is real detail, what its track contains, and whether an audio reference ever drives the lips.
Whether a start frame and references can be attached together on the connection.
Whether Gemini Omni Flash 1.1 should replace Veo 3.1 as the sound-led route once its contract on the connection is read (Chapter 20).