Guide
MiniMax H3 Max Image-to-Video Prompt Recipes for Developers
Copy-paste prompt structures, camera-motion wording, and negative lists you can test against your own first frames
On this page
Why prompt structure beats prompt length
Most disappointing image-to-video results don't come from a prompt that was too short. They come from a prompt that was long, vague, and internally inconsistent — a wall of cinematic adjectives that never tells the model what to actually keep and what to actually move.
With image-to-video, the first frame already carries most of the visual information: identity, framing, lighting, palette, texture. Your prompt's job is narrower than you think. It should (1) confirm what must stay, (2) specify what should change, and (3) describe how the camera behaves while that change happens. Adjectives like "epic" or "breathtaking" don't map to any of those three jobs.
The recipe skeleton
Every recipe in this post uses the same six blocks, in the same order:
- Subject — who or what is in frame, with the identifying details that must survive.
- First-frame state — what the still shows: pose, framing, light direction, background layers.
- Motion — the single primary movement in the scene, plus any secondary motion.
- Camera — one movement type, one direction, one rough magnitude.
- Pacing — shot length and whether the motion resolves or continues.
- Negatives — the artifacts you want suppressed.
Writing in this order isn't cosmetic. It mirrors how the conditioning actually applies: you anchor the still, then describe deltas. When a recipe fails, a fixed skeleton lets you point at one block and change it, instead of rewriting the paragraph and guessing which sentence broke it.
Why a skeleton makes recipes comparable
If every prompt is free-form prose, you can never tell whether a better output came from your new camera phrasing or from the fact that you also happened to reword the subject. A skeleton converts prompt tuning from an artisanal craft into a controlled experiment: same subject, same first-frame state, one changed block.
Keep each block short — a clause, not a paragraph. The goal is a prompt you can diff line by line in code review.
The first-frame description block
The first-frame block exists to fight drift. Image-to-video models are conditioned on the input still, but a prompt that ignores the still — or worse, contradicts it — invites the model to re-render your subject into something generic. Your description should read like an annotation of the frame you actually uploaded, not a wish list for a different frame.
Annotate, don't re-imagine
Cover four things, briefly:
- Subject identity: the details that must not change — hair length and color, a jacket's cut, a product's label placement, a building's window rhythm.
- Framing: close-up, medium, wide; subject position in frame ("centered," "left third"); how much headroom or floor is visible.
- Light direction and quality: "soft key from camera-left," "hard backlight with rim on the shoulders," "overcast, no visible shadow."
- Background layers: what sits behind the subject and how far away it reads — "blurred city street," "flat seamless backdrop," "treeline roughly ten meters back."
Patterns that work
Subject: a woman in her thirties, shoulder-length dark hair, olive field jacket, no glasses.
First frame: medium close-up, centered, chin at center height, soft key light from camera-left, cool fill on the right cheek, blurred storefronts behind her at street distance.
Notice what's absent: no mood adjectives, no backstory, no color grading instructions that fight the still's existing grade. You are describing, not directing a remake.
Common failure: over-describing
If your first-frame block runs longer than your motion block, you've inverted the priorities. Long descriptions invite contradictions — you mention "bright daylight" while the still is dusk, and the model has to choose. When the still and the text disagree, drift is the usual outcome. Write the block by looking at the frame, then delete anything the frame already makes obvious.
One practical habit: keep the reference frame and the first-frame block in the same file, adjacent. If you edit one, re-read the other.
Camera motion and timing phrasing
Vague cinematic language is the most common source of "random" motion. "Dynamic camera work" tells the model nothing actionable. Name the movement, give it a direction, give it a magnitude, and say how long it runs.
Wording patterns for the four workhorse moves
Push-in / dolly in
Camera: slow push-in toward the subject's face, roughly 15% closer over the shot, no rotation.
Pan
Camera: steady horizontal pan left to right, moderate speed, horizon stays level.
Orbit
Camera: slow orbit around the subject, about 30 degrees to camera-right, subject stays centered, background parallax visible.
Static with subject motion
Camera: locked off, no movement. Only the subject moves.
Each of these does three things: one movement type, one direction, one rough magnitude. Adding a second movement ("push in while orbiting") is legal but should be deliberate — it doubles the chance of an unintended result.
Timing and pacing
Say whether the motion resolves or continues, and anchor it to shot length in plain words:
- "the push-in completes by the final frame"
- "the pan continues past the end of the shot"
- "hold static for the first half, then begin the orbit"
- "single continuous take, no cuts"
Avoid numeric second counts unless your pipeline fixes the clip duration; if it does, matching your wording to that duration removes an entire class of failure. "No cuts" is worth stating explicitly, since some prompts otherwise invite an implied edit.
Speed words that actually mean something
Replace vague terms with relative ones the model can calibrate against: "slow," "moderate," "fast," and "almost imperceptible." These are far more useful than "cinematic," "sweeping," or "dramatic," which carry no direction or magnitude. A good test: could a camera operator read your line and know which way to move? If not, rewrite it.
Negative lists by content type
Negative terms are cheap insurance, but they're not universal. A negative list tuned for talking-head footage can actively harm a product turntable. Keep separate lists per content type and only extend a list when you see a repeatable artifact.
People
Negatives: extra fingers, deformed hands, warped facial features, identity drift between frames, flickering eyes, uncanny teeth, duplicated limbs, clothing texture shimmer
Identity drift and hand artifacts are the two chronic issues. If your subject turns or gestures, add "no morphing of facial structure during head turn."
Products
Negatives: label text warping, changing logo shape, surface reflections flickering, geometry wobble on straight edges, color shift across frames, background objects appearing
Product shots fail on consistency, not drama. Straight edges and printed text are the canaries — if the label breathes, the shot is unusable.
Environments
Negatives: objects popping in or out, foliage jitter, building lines bending, water surface pulsing, sudden lighting changes, horizon tilt during pan
For pans especially, "horizon tilt" and "warping of straight architectural lines" earn their place in the list.
Extend versus trim
Add a negative when you can reproduce the artifact across at least two runs with the same frame and prompt. Remove negatives you've never seen fire, especially when they conflict with your intent — a "no camera movement" negative in a push-in recipe is a contradiction that costs you stability. Keep each list under a dozen items; a long negative list dilutes the guidance just like a long positive prompt does.
Store the list as a named string in your config, not pasted inline per prompt, so all recipes of a content type stay aligned.
Recipe walkthrough with real outputs
The recipes below are complete and copy-pasteable. Descriptions of the outputs are from test runs against our own first frames; treat them as expectations to verify, not guarantees, since your frames differ.
Recipe 1: portrait push-in
Subject: a man in his forties, short grey-flecked hair, navy crew-neck sweater.
First frame: medium close-up, centered, soft key from camera-left, warm practical light in the blurred background at room distance.
Motion: subtle head turn toward camera, slight shoulder settle; no walking.
Camera: slow push-in toward the face, about 15% closer over the shot, no rotation.
Pacing: single continuous take, push completes by the final frame.
Negatives: extra fingers, warped facial features, identity drift between frames, flickering eyes, sudden lighting changes.
Output: the identity held, and the push read as intentional rather than a zoom crop. Earlier versions without the "no rotation" clause produced a slight roll that made the frame feel handheld.
Recipe 2: product turntable
Subject: a matte ceramic mug with a printed label on one side.
First frame: product centered on a seamless light-grey backdrop, soft top light, faint contact shadow.
Motion: none from the product; the label becomes readable as the view changes.
Camera: slow orbit to camera-right, about 30 degrees, subject stays centered.
Pacing: continuous, orbit continues past the end of the shot.
Negatives: label text warping, changing logo shape, geometry wobble on straight edges, color shift across frames.
Output: the label stayed legible through most of the arc; the last third showed slight edge softening, which we fixed by reducing the orbit to roughly 25 degrees rather than adding more negatives.
Recipe 3: environment pan
Subject: an empty coastal road at overcast midday, low guardrail on the sea side.
First frame: wide shot, road receding to the right, flat diffuse light, no visible sun.
Motion: gentle wind in roadside grass; nothing else moves.
Camera: steady horizontal pan left to right, moderate speed, horizon level.
Pacing: continuous pan, no cuts.
Negatives: horizon tilt during pan, building lines bending, objects popping in, water surface pulsing.
Output: the horizon stayed level because we named it. Without that clause, earlier runs drifted a few degrees over the shot — subtle, but enough to feel wrong on a wide landscape.
Iterating and regression-testing prompts
A prompt is not a one-off message; it's a configuration value. Treat it the way you treat any other config that affects output: version it, log what it produced, and change one variable at a time.
The one-variable loop
- Pick a failing behavior in the output (drift, hand artifact, unwanted camera roll).
- Identify which block owns it — usually motion, camera, or negatives.
- Change exactly one clause in that block.
- Re-run the same first frame.
- Keep or revert based on what you see, then log it.
Changing two blocks at once is how teams end up with prompts nobody can reason about six weeks later.
Log frame, prompt, and output together
Every run record should contain:
- a hash or ID of the input frame
- the full prompt string, including negatives
- the recipe version
- the output reference and a short note on what passed or failed
This triplet — frame, prompt, output — is what makes regressions detectable. When the underlying model updates, or when you tweak a shared negative list, you can re-run your stored frames and diff the results against known-good outputs.
Build a small golden set
Keep five to ten frames per content type that represent your actual workload: a portrait with turned head, a product with printed text, a wide environment with a horizon. Run every recipe change against this set before it reaches production. It takes minutes and catches the majority of accidental regressions.
Track expectations, not just scores
A simple pass/fail per case — "hands clean," "label legible," "horizon level" — is more useful than a single subjective quality score, because it tells you which block to revisit. When a case flips from pass to fail, you know exactly which recipe version and which clause to blame.
Wiring recipes into your pipeline
Once recipes are stable, they should stop living in a notes app and start living in code.
Store recipes as versioned config
Keep each recipe as a structured object — one field per block — rather than a pre-concatenated string. Joining blocks at request time lets you share subject and first-frame blocks across recipes, and swap negative lists per content type. Put the config in version control, tag it, and include the version in every log record.
{
"id": "portrait_pushin",
"version": "3",
"subject": "...",
"first_frame": "...",
"motion": "...",
"camera": "slow push-in toward the face, about 15% closer, no rotation",
"pacing": "single continuous take, push completes by the final frame",
"negatives": "portrait_default"
}
Batch-test before you ship
Run the assembled prompt across your golden frames in a loop, store outputs alongside the run record, and review flagged cases before promoting a recipe version. Because every recipe hits the same endpoint through the same interface, a batch harness is a thin wrapper around your normal call — no per-model special cases to maintain.
Call the endpoint at scale
In production, assemble the prompt from config, pass your first frame, and submit through the API. Treat rate limits and retries the same as any other async workload: queue jobs, attach the recipe version and frame ID to each job, and write the returned output reference back to your log. If a recipe misbehaves in production, you can pull the exact frame and prompt that produced it.
For endpoint parameters, supported aspect ratios, and request/response specifics, see the MiniMax H3 Max image-to-video model page — the recipes above are designed to drop straight into it.
Write your prompts like code and your outputs like test results, and image-to-video stops being a slot machine.
Related articles
View all articles →-
Guide
gpt-6.2-sol Getting Started Guide: Capability Boundaries, API Integration, and Token Pricing Essentials
Integration decisions and practical essentials for engineering teams
-
Guide
gpt-6.2-astra Getting Started Guide: Text Model Capability Boundaries, Integration Essentials, and Cost Control
-
Guide
Tripo H3.1 Getting Started Guide: Choosing Among Three Input Modes, Parameters, and Pricing
A practical guide for developers on 3D generation integration decisions and cost control