Guide
Flux-3-Max Image-to-Video Prompt Recipes: Copy-Paste Templates for Developers
First-frame description structure, motion and camera wording, and stability negatives you can reuse
On this page
Why prompt structure matters more than prompt length
Developers new to image-to-video often write prompts like a screenplay: a long paragraph describing the character's backstory, the mood, the color palette, and a dozen things that should happen. That approach fails because the model is not reading a narrative — it is conditioning a short generation on a source frame. The more prose you add, the more ambiguous the signal becomes.
What works instead is ordered structure. A prompt for flux-3-max/image-to-video should read like a small config file: first anchor the frame, then declare motion, then constrain the output. Each block answers one question.
- First-frame block — what is already visible in the source image? Subject, scene, lighting.
- Motion block — what changes over the clip? Subject action and camera movement.
- Negative block — what must not appear? Artifact classes only.
This ordering matters because the model attends to the source frame first. If the opening words describe a scene that doesn't match the image, everything downstream drifts. If the opening words confirm the frame, the model has a stable base for the motion instructions that follow.
Length is not the enemy by itself — redundancy is. Ten words that each do a distinct job beat a hundred words that restate each other. When you review a prompt, ask of every phrase: does this change what the model renders? If not, delete it. The templates later in this guide are short for exactly this reason.
Keep one more rule in mind: a prompt is not a wishlist. Stacking competing instructions ("slow push-in while orbiting and zooming out") gives the model contradictory physics to resolve, and the result is usually jitter rather than a smooth move. Single-intent phrasing is the theme that runs through every section below.
The first-frame description block
The first-frame block is a description of what the source image already contains, rewritten in the model's language. You are not inventing a scene — you are telling the model what it is looking at so the motion instructions have a subject to act on.
Subject
Name the main subject with one noun phrase and, if useful, one material or state adjective: a ceramic mug on a desk, a woman in a linen shirt, a mountain ridge at dusk. Avoid adjectives that describe mood — those belong nowhere in a prompt, because they don't map to pixels.
Scene
State the setting in one clause: in a bright studio, against a concrete wall, on a forest trail. This helps the model decide how much of the background it should keep stable versus allow to shift.
Lighting
Lighting is the most under-specified part of most prompts and the most visible in output. One short clause is enough: soft window light from the left, overcast daylight, warm lamp glow. Naming the direction of light, not just its quality, gives the model a consistent shading assumption across frames.
What not to include
The first-frame block is not the place for backstory, emotion, or future events. If you want the subject to smile, that is a motion instruction, not a description. Keeping the block purely descriptive prevents the model from hallucinating details the frame doesn't support.
A complete first-frame block might read: a ceramic mug on a wooden desk, in a bright studio, soft window light from the left. That is the entire anchor. Everything after it is motion.
Wording for motion and camera movement
Motion wording works best when it is physical, short, and single-intent. The model responds to verbs of movement and to camera geometry, not to directorial abstraction.
Subject motion
Use one action clause: the steam rises slowly, she turns her head slightly, the fabric shifts in the wind. Keep the subject and the verb close together. Avoid chaining two actions on the same subject unless they are naturally linked — he lifts the cup and drinks is fine; he lifts the cup while walking away is not.
Camera movement
Camera moves are best expressed as a short phrase naming the operation and a direction: slow push-in, pan left to right, orbit around the subject, slight tilt up. The word slow or slight is often more useful than a numeric speed, because it sets intent without demanding an exact rate the model cannot reliably honor.
Pick one camera move per clip. Stacking moves is the single most common cause of unstable output: when a push-in and an orbit are requested together, the model tries to satisfy both and produces warping around the subject's edges. If you genuinely need a compound move, split it across two clips or accept that the second move will be approximate.
Separating subject and camera
Write subject motion and camera movement as two separate sentences. The steam rises slowly. Slow push-in. This makes it easy to change one without touching the other — which, as the iteration section explains, is how you debug a prompt.
Phrases to avoid
Avoid abstract direction: cinematic, epic, dynamic, emotional. These don't specify a pixel change, so they add noise. If you want a specific look, describe the camera geometry or the light, not the feeling.
Stability negatives that actually help
Negative prompts are frequently misused as a second style description — a place to list everything you dislike. That over-constrains the model and often makes output worse, because the generation is now avoiding many things at once instead of doing one thing well.
The productive use of negatives in flux-3-max/image-to-video is artifact suppression. Narrow, technical negatives tell the model which failure modes to avoid without touching the creative direction.
Useful artifact negatives
warping,morphing,flickering— temporal stability issues.extra limbs,duplicate faces— structural errors around people.text artifacts,watermark— unwanted glyph-like detail.blurry edges,jagged silhouettes— boundary instability.
These are all descriptions of how generation fails, not of how you want it to look. That is the test for whether a negative belongs.
Negatives that usually hurt
Avoid negative style words: no cinematic look, not cartoon, no bright colors. These fight the positive prompt and can flatten the output. If the positive prompt is well-specified, style negatives are unnecessary.
Keep the list short
Two to five narrow negatives are usually enough. A long negative list increases the chance of conflicting constraints, and the model may suppress legitimate detail along with the artifact. Start with the Temporal stability and boundary items, add others only when you observe the corresponding artifact in a test run.
Copy-paste recipe templates
These skeletons follow the order from the earlier sections: first-frame block, then motion, then negatives. Replace the bracketed slots and keep the structure intact.
Product clip
[product noun phrase] on [surface], in [setting], [light direction and quality].
[one subtle subject motion, e.g. reflection shifts / light passes across the surface].
[single camera move, e.g. slow push-in / slight orbit].
Negative: warping, flickering, text artifacts.
Portrait clip
[person noun phrase] in [clothing], against [background], [light direction and quality].
[one small subject motion, e.g. head turns slightly / eyes blink].
[single camera move, e.g. slow push-in / slight tilt up].
Negative: warping, extra limbs, duplicate faces.
Environment clip
[landscape or interior noun phrase], [time of day and weather], [light direction and quality].
[one ambient motion, e.g. clouds drift / leaves move in the wind].
[single camera move, e.g. slow pan left to right / slight tilt up].
Negative: flickering, blurry edges, morphing.
Adaptation rules
- Keep the subject phrase to one noun and at most one modifier.
- Change only the motion verb when you want a different action — leave structure fixed.
- If a clip looks unstable, remove the camera move before changing anything else; the subject motion alone is usually the safer baseline.
These templates are starting points, not final prompts. The source frame interacts with phrasing more than most developers expect, so treat every template as a hypothesis to validate on one short run before committing it to a batch.
Iterating and batching across clips
The fastest way to converge on a good prompt is to change one variable at a time and inspect a short run. A practical loop looks like this:
- Write a prompt from the closest template.
- Generate a short clip.
- Inspect it against the source frame — is the subject preserved? Is the motion single-intent? Are artifacts present?
- Identify the single block responsible and change only that block.
- Regenerate and compare.
This loop is deliberately slow on the first clip and fast afterwards, because each fixed block becomes a reusable asset. Most instability traces back to either a stacked camera move (motion block) or a subject phrase that doesn't match the frame (first-frame block). Changing one at a time tells you which.
Scaling to a batch
Once a prompt is stable on one frame, promote it to a template and apply the same structure across similar source frames. Batch iteration works because the structure is constant and only the slots change:
- Keep the first-frame block in sync with each new image; do not copy a description from a different frame.
- Keep motion wording identical unless the clip type changes.
- Keep the negative list fixed per clip category.
Review batches in groups by clip type, not by individual file, so a systematic problem (for example, all portrait clips warping at the shoulders) is visible immediately rather than buried in per-file notes.
Finally, record which prompt variant produced which output. A short changelog of prompt edits per clip category will save more time than any single clever phrase. Build the loop, keep the structure, and let the templates carry the repetition.
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