Guide
Hailuo 2.3 Fast Image-to-Video Prompts: Copy-Paste Recipes for Developers
First-frame structure, motion phrasing, quality trade-offs, and negative prompts for the fast tier
On this page
What the fast tier is good for
Fast tiers earn their place in a workflow for one reason: how many attempts you can afford. Image-to-video prompting is an empirical process — you form a hypothesis about how a still frame should move, generate, look at the result, and adjust one thing. A faster model compresses the feedback loop, which means you find the good prompt sooner even if any single clip is less polished than what a slower tier would return.
So reach for hailuo-2.3-fast when the prompt itself is what you are still figuring out:
- Prompt discovery. You have a first frame and no idea how the model will interpret it. Run cheap, fast variations until the motion vocabulary clicks.
- Composition checks. Does the camera anchor hold? Does the subject stay in frame? These are structural questions that do not need maximum fidelity to answer.
- Iteration on one variable. Speed of a gesture, direction of a turn, whether a prop moves — each is a single change best tested in sequence.
- Batch sanity passes. Ten short attempts across different phrasings tell you more than one long attempt at a single phrasing.
Reach for a slower tier when the recipe is already settled and you are producing on a specific shot — where texture, fine detail, and long-range coherence matter more than how many tries you get. The practical pattern is to treat the fast tier as your staging environment and a slower tier as production.
One structural note: on the fast tier, prompt budget is not infinite. Short, ordered descriptions survive better than sprawling paragraphs, which is the subject of the next section.

A repeatable first-frame description structure
The input image already encodes most of the visual information. The prompt's job is not to re-describe the picture — it is to tell the model which parts of the picture are load-bearing and what should change. A fixed ordering makes runs comparable: if you always describe things in the same order, a difference between two outputs is more likely attributable to the one thing you changed.
Use four blocks, in this order.
1. Subject
Who or what the frame is about, and any state that matters for motion. Keep it to the salient actor plus one or two attributes that constrain movement — posture, what the hands are holding, whether the figure is seated or standing.
Subject: a woman in a canvas jacket, seated, hands resting on a wooden table
2. Setting
Enough environment to establish what can move and what is fixed. Name surfaces, light direction, and any element that should react to the subject (steam, fabric, foliage, reflections).
Setting: narrow cafe interior, window light from the left, background softly out of focus
3. Camera anchor
State explicitly whether the camera is locked off or moving. Ambiguity here is the most common cause of unwanted drift on short clips.
Camera: locked-off medium shot, no pan, no zoom
4. Motion cue
What happens, in one sentence, with a direction and an endpoint. Details in the next section.
Motion: she lifts the cup to her mouth and sets it back down on the table
Assembled, the prompt reads as one compact paragraph: subject, setting, camera, motion. Resist the urge to add texture adjectives in the first pass — get the motion right, then add detail selectively, which is exactly the trade-off the fast tier forces.

Motion phrasing that behaves like physics
Abstract motion words are the biggest source of unstable clips. "Dynamic," "cinematic movement," "energetic" and similar adjectives describe a feeling, not a trajectory, and the model has to guess what you meant. Concrete phrasing that names a direction, a speed, and a sense of weight produces clips that stay coherent because every clause constrains the outcome.
Three properties carry most of the load:
- Direction — which way, relative to the frame. "Toward the camera," "left to right across frame," "downward into the lower third." If it matters, say it.
- Speed — slow, steady, quick, abrupt. A single adverb changes how many frames the action spans, which changes whether it finishes inside the clip.
- Weight — what resists and what gives. Heavy things start slowly and stop slowly; light things snap. "The curtain settles" reads differently from "the curtain flaps."
Compare:
Weak: the scene comes alive with movement
Better: the leaves shift gently left to right in the breeze
Weak: she moves dramatically
Better: she turns her head slowly to the right, then holds still
Two more habits worth adopting. First, bound the action: say where the motion ends, or that the subject returns to rest. Clips drift when the model keeps generating motion past the point you described. Second, sequence when there is more than one action: "lifts the cup, then sets it down" gives an order; "a cup lifting and a hand moving" does not.
Keep the whole motion clause to one or two sentences. Long chains of actions split the model's attention, and on the fast tier they are the first thing to fall apart.
Copy-paste recipes
Each template below is a complete prompt skeleton. Replace the bracketed slots and keep the ordering intact. The bracketed variables are the only things you should change on a first pass — change one at a time.
Recipe 1: Locked-off portrait with a single gesture
Subject: [a person], [posture], [what the hands are doing]
Setting: [interior/exterior], [light direction], background softly out of focus
Camera: locked-off medium shot, no pan, no zoom
Motion: [subject] [single gesture with direction], then holds still
Negative: no camera shake, no warping, no extra fingers
Use this when the first frame is a clean portrait and you want lifelike micro-movement without the camera doing anything. The variable that matters most is the gesture clause — swap it first.
Recipe 2: Environment reacting to a still subject
Subject: [figure or object], motionless, [posture or state]
Setting: [environment] with [moving element: steam, fabric, foliage, water], light from [direction]
Camera: locked-off wide shot, no pan
Motion: [moving element] drifts [direction] at [slow/steady] speed; subject remains still
Negative: no subject movement, no camera drift, no flickering
Use this when you want atmosphere rather than action. Naming the still subject as motionless in both the subject and motion blocks is what keeps the figure from drifting.
Recipe 3: Product or object turn
Subject: [object] on [surface], [notable material or finish]
Setting: [studio/surface], [key light direction], clean background
Camera: locked-off close-up, no zoom
Motion: [object] rotates [clockwise/counter-clockwise] [slowly/steadily] until [end position]
Negative: no shape warping, no background movement, no text artifacts
Use this for inspection-style clips. The end position matters: an unbounded rotation tends to keep spinning past what looks natural in a short clip.
Adapting without breaking the structure
When a recipe does not land, resist rewriting everything. Identify which of the four blocks the failure belongs to — wrong motion, unwanted camera move, wrong setting emphasis — and edit only that block. The notes in each recipe flag the highest-leverage variable so you know where to start.

Trading off quality on the fast tier
Fast generation settings leave less room for the model to resolve fine detail per frame, so prompt tokens spent on texture are tokens not spent on structure. The practical rule: on the fast tier, spend prompt budget on motion and composition clarity, not surface detail.
Where detail pays off:
- Motion endpoints and direction — the single highest-value words in the prompt. A clearly bounded action survives fast settings far better than a vaguely described one.
- Camera statements — an explicit locked-off or explicit pan prevents drift, which is more visible than missing texture.
- Which element moves — naming one moving element and one still element is worth more than five adjectives.
Where detail is wasted:
- Fabric weave, pore-level skin texture, individual hair strands. The first frame already carries this; re-describing it does not add resolution the fast tier is not going to render anyway.
- Stacked quality adjectives. "Photorealistic, 8k, ultra-detailed" is noise. Composition and motion are the levers.
- Long lists of style references. They compete with the motion clause for attention.
A useful test: strip your prompt down to subject, camera, and motion. If the clip is still wrong in structure, no amount of added detail will fix it. If the structure is right but the result feels flat, add back one or two setting details and re-run. That is the whole fast-tier quality workflow — build structure first, decorate second, and only decorate what you can see the model responding to.
Negative prompts that actually help
Long generic negative lists are mostly cargo cult. Copying fifty terms from someone else's config does not tell the model anything about your image, and terms that do not correspond to a real failure mode just consume attention. Effective negative prompts are short, concrete, and derived from artifacts you actually observed.
A workable loop:
- Generate with no negative prompt.
- Watch for the two or three specific things that went wrong.
- Name those things plainly in the negative field.
- Re-run with the same positive prompt plus the negative additions.
- Stop when the visible artifacts are gone — do not keep padding the list.
Common artifact names, phrased plainly:
Negative: no camera shake, no warping of [the object], no extra fingers, no background drift
Keep the wording matched to the artifact. "No camera shake" is better than "no instability." "No extra fingers" is better than "no anatomical errors." Concrete nouns and verbs give the model a specific thing to avoid.
Two cautions. First, negatives that contradict your positive prompt waste the run — if you asked the subject to turn, do not include a blanket "no subject movement." Second, once a negative has fixed a problem, keep it; treat the negative field as an accumulating, per-image record of known failures, not a scratchpad you rewrite from memory each time.
Testing prompts like code
Treat the prompt as source and the output as a test result. The minimum viable version of this costs almost nothing in tooling:
- Version every prompt. A short id in a comment or filename, plus the input image reference, is enough. You need to be able to answer "what exactly produced this clip?" a week later.
- Change one variable per run. Different image or different motion clause — never both in the same comparison. If you change two things and the clip improves, you have learned nothing about which one mattered.
- Log the artifact, not just the verdict. "Clip 7: subject's left hand drifted out of frame" is useful. "Bad" is not.
- Keep a negative-prompt ledger per image. Once an artifact is fixed, the negative term that fixed it belongs to that image's recipe.
- Promote recipes, not one-offs. When a template lands reliably across two or three different images, it graduates into your reusable set.
A small structured note per run — prompt id, image id, one-line change, one-line observed artifact — turns prompt work into something you can reason about instead of guess at. Over a dozen runs on the fast tier, that log is the difference between having a prompt and having a process.
Start with one image you know well, run the three recipes against it, and let the log tell you which motion phrasing your subject actually responds to.
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