Same prompt, three models: what actually happens

A useful test when picking a platform: run the same 1,500-character prompt on all three. Example — "A Yoruba woman in a black gele turns slowly toward the lens in a brown void. 35mm anamorphic, soft directional key from upper-left. Painterly hand-drawn aesthetic with visible brushwork."

• Seedance: continuous motion, strong atmospheric register, painterly stylization decent but inconsistent. Most cinematic feel.
• Runway: cleaner motion, less continuity, weakest painterly stylization of the three. Most "AI-cinematic" looking.
• Dreamina: strongest painterly look, motion slightly choppy, most "illustrated" feel.

The platform IS the choice. Pick based on the deliverable.

▶ Fashion Film — AI-directed fashion motion: the continuous fabric behaviour and unbroken movement that expose how differently each model handles motion from the same brief

Which AI video model should you use?

The decision tree, in order:

1. Multi-beat sustained piece with atmospheric continuity? → Seedance.
2. Single shot or a product cinematic? → Runway.
3. Painterly / illustrated aesthetic? → Dreamina.
4. Needs to read as found footage? → Runway (with handheld discipline).
5. Needs a cosmic-scale environment? → Seedance.
6. Needs a packshot? → Runway.
7. Needs cultural specificity in styling? → Seedance or Dreamina, depending on register.

And when a piece is failing, diagnose platform vs prompt before you rewrite:

• Motion feels generic across all platforms → prompt issue; your motion specification isn't specific enough.
• Motion works on Seedance but fails on Runway → platform issue; use Seedance.
• Atmosphere works on Seedance but breaks on Dreamina → wrong tool for the job.
• Identity drifts across all platforms → your identity block needs work: more specific, locked earlier in the prompt.

AI video prompt character limits (and the sub-cap sweet spot)

The hard caps: Runway 3,500 characters. Seedance 4,000. Dreamina 4,000.

But do not write to the cap — write 50–100 characters under it. The model performs best with a small margin, and the buffer avoids edge-case rejections from platforms that count whitespace slightly differently. Targets:

• Runway: 3,450–3,499 characters
• Seedance: 3,950–3,999 characters
• Dreamina: 3,950–3,999 characters

Validate with the character-count workflow before shipping:

cat << 'PROMPT_END' | wc -c
[paste prompt here]
PROMPT_END

This counts characters including newlines. Match against your platform cap (Runway 3,500, Seedance 4,000).

Each platform also compresses differently:

• Seedance: cut decoration words first ("beautiful", "stunning", "epic"); keep technical specificity; aim for 3,950–4,000 — the model uses every available character.
• Runway: cut clauses first; each beat should be one or two sentences; aim for 3,300–3,500; use a closing constraint block for last-resort discipline.
• Dreamina: cut atmospheric continuity, increase aesthetic specificity; aim for 3,800–4,000 — the model rewards style words more than continuity.

How to compress an AI video prompt without killing it

You will hit the character cap. Constantly. Every word in a prompt has a job; every word without a job is a cost. The hierarchy of what to cut, in order:

1. Decoration words (beautiful, stunning, gorgeous, magnificent, breathtaking, epic, cinematic, atmospheric)
2. Hedging words (somewhat, slightly, a bit, kind of, perhaps)
3. Redundant adjectives ("the slow, gentle, gradual push" — pick one)
4. Restated content (the same thing said twice in different ways)
5. Atmospheric repetition (mood register repeated across sections)
6. Optional supporting details (the third secondary motion, the fourth ambient sound)
7. Final-beat decoration

Cut in this order. Cut hard. The piece survives.

Semantic compression

Retain meaning, reduce characters: "extremely large" → "vast"; "very slowly" → "languidly"; "in order to" → "to"; "the woman who is dancing" → "the dancing woman"; "she is standing in the middle of" → "she stands in"; "located at" → "at". These small surgeries add up — a 4,200-character prompt becomes 3,900 without losing a single image.

Replace conceptual words with visual words

The highest-leverage compression technique. Concepts are vague; visuals are specific:

• "Cinematic" → "35mm anamorphic, film grain, lifted blacks"
• "Atmospheric" → "volumetric haze, soft particle drift"
• "Beautiful lighting" → name the source, temperature, quality, and behaviour
• "Epic" → "wide cosmic, deep void, sub-bass under everything"

The visual replacements are longer in characters, but the model uses them better. Better images per character is the actual metric. The compression theorem: sometimes longer is shorter — a longer specific phrase that generates correctly is more efficient than a shorter vague phrase that requires regeneration. Same logic for emotion: "She is full of grief" (vague) loses to "Her eyes lower. Her hand falls" (specific).

What to protect

When cutting, preserve motion specification and cut atmospheric flavour first. Camera motion is sacred; character gesture is near-sacred; secondary motion (fabric, earrings) goes first under pressure, then environmental ambient; VFX accents go before all of them. If you must cut a whole beat, cut a transitional one — almost never the opening or closing beat.

The lines you never trim, no matter what:

1. The identity lock (e.g. "Identity locked, zero drift.")
2. The camera global instruction (the dominant move)
3. The opening beat's first sentence
4. The closing beat's last sentence
5. The thesis line
6. Filter-trigger replacements (e.g. "premium dark spiced rum bottle" instead of "Captain Morgan")
7. Colour discipline percentages
8. The "NOT" lines in the style block (NOT flat black, NOT modern luxury gold, NOT photorealism)

Then re-read once before shipping: camera motion clear in every beat? Identity locked? Thesis line intact? Negatives strong? The most common compression failure is stripping a protected line that looked like decoration — the re-read catches this 70% of the time.

Porting prompts between platforms

Standardising a stack means moving prompts between tools. Seedance → Runway:

1. Trim to 3,500 characters or less — cut atmospheric detail first, then beat-level detail, then identity detail (preserve identity longest).
2. Strip filter triggers — brand names, named cinematographers, named photographers, specific products; replace with descriptive equivalents.
3. Reduce beat count — Runway handles fewer beats better; 3 beats max in 10 seconds.
4. Increase verb directness — Runway responds to direct verbs; Seedance tolerates clause-heavy description.
5. Re-test the closing thesis line — Runway sometimes ignores the final sentence; if it's critical, move it higher.

Seedance → Dreamina: emphasize stylization earlier, reduce camera technical language slightly, add explicit material descriptors. Runway → Seedance: add atmospheric continuity, held-frame instructions and a motion hierarchy — Runway prompts tend to be too sparse for Seedance to render well.

Motion psychology: why some AI video feels alive

Motion is meaning. Every camera move carries emotional payload — but models don't understand this, so you specify both the motion and its emotional intent. The core vocabulary:

• Slow inward push = emotional gravity. Name the arrival: "Slow forward push, gently closing distance, ending at intimate range" beats "camera dolly forward."
• Slow pullback = revelation. Specify the scale it reveals: "Final frame: subject as a tiny island in vast emptiness."
• Orbit = mythic perception. The subject reads as a monument. One orbit per piece — multiple orbits become a music video.
• Locked frame = sacred witness. Specify what moves within the lock: wind, fabric, light, breath.
• Micro-fabric movement = life presence. "Her robe drifts" is generic; "Her robe drifts, edge catching the gold reflection as she breathes" is alive.
• Asymmetry = realism. "Slightly rotated chairs, compressed cushions, uneven object placement" turns a generic interior into a lived-in space.
• Imperfect timing = authenticity. "Frame drifts slightly with breath", "focus hunts briefly before finding" — imperfection reads as truth.
• Environmental lag = physicality. "ONE BEAT of silence. Then the water rises." Instant effects read as cartoon; lagged effects read as physics.
• Delayed secondary motion = weight. "Her head turns slowly toward camera-left. Half a beat later, her gold earrings swing once."
• Camera breath = presence. "Subtle anamorphic breathing" costs one phrase and turns a tripod into a body.
• Held gaze = connection. Sweet spot: 2–4 seconds in a 15-second piece, at a specified timestamp: "her eyes find the lens at 0:03 and HOLD."

Motion is punctuation: constant motion is run-on prose, no motion is fragmented. Big motions earn their amplitude by following held moments. And the camera's position is an emotional position — eye level is trust, low angle makes the subject monumental, bird's eye is detached. Pick for the emotional experience you want, not just the framing.

The single best motion specification from this system: "Camera locked. Everything moves ever so slightly — panels micro-correcting, drones drifting, ship lanes flowing — but the camera does not." One sentence: camera state, three named micro-motions, the principle, and an implied reverence. The move for monumental calm.

Negative prompting: the discipline of saying NO

Models default to the most generic version of any concept. To specify what you want, you must also specify what you do not want. The negative prompt is not a fallback — it is a structural tool. The most powerful negative in this system is three words: "NOT flat black." Applied to dark skin, it changed the render from a featureless dark patch to warm reflective ebony with sculptural detail. One instruction, massive change.

The sanity-check baseline

Include these almost universally — they prevent the most common AI video failures:

• "No CGI gloss."
• "No photorealism unless specified."
• "No saturation drift."
• "No identity drift."
• "No fast motion unless specified."

Then add piece-specific anti-patterns. For painterly pieces: "NO photorealism. NO smooth animation." For found footage: "NO impossible angles. NO Hollywood coverage. Camera always makes mistakes." For cosmic pieces: "Documentary stellar physics. NO cyberpunk. NO neon. NO floating UI." For character work: "Identity locked, zero drift. NO facial drift. NO costume drift. NO multiple versions of subject."

▶ Miriam — The Keeper of Thresholds — a character film held together by exactly these negatives: identity locked across every beat so the emotion, not the drift, is what you watch

Placement and phrasing rules

• Put the constraint block LAST. The last thing the model reads weighs heavily on its final state. Constraints placed first get respected early and drift in the middle; constraints at the end produce more consistent generations.
• Rewrite negatives as affirmations where stronger. "No saturation drift" → "Colour values held constant across all 15 seconds." "No identity drift" → "Same face, same hair, same jewellery from beat 1 to beat 5." Use both — the redundancy reinforces the constraint.
• Use negatives as a filter-safe zone on Runway. Instead of "Captain Morgan" (filter trigger): "premium dark spiced rum bottle, red-and-gold label" in positive, "NOT generic alcohol product" in negative. Renders the right thing without the flag.
• Don't negate what the model wouldn't generate anyway. "No horses in frame" in a space piece wastes characters — and the model can read the negation as a hint. Negate only real defaults: generic cinematic look, saturation drift, floating particles, AI smoothness.
• Force one register. "One painterly register only. No style mixing." When a piece has multiple stylistic references, models try to mix them all — this instruction forces a commitment and has saved more generations than any other negative.

Learn this live

This is Part 2 of Director's Prompt: The AI Secret Sauce — the published series built from the Seedance Prompting Bible working manual — and it's exactly what we teach hands-on at PxLabs in Lagos. If you want to direct AI video like a cinematographer rather than gamble with prompts:

• Take the course: Generative Cinematography — live AI training in Nigeria covering model selection, prompt architecture, compression and negative prompting on real briefs.
• Join the community: PxLabs Community — creators, marketing teams and agencies comparing notes on Seedance, Runway and Dreamina workflows.
• Need ads made? As an AI agency in Lagos, we produce AI video ads for brands end-to-end. Talk to us.

Next in the series

Part 3 covers the prompt templates and keyword library — the reusable block structures, camera vocabulary and style keywords that turn everything above into copy-paste production assets.