One Photo, Six Shots: How This 15-Second Seedance 2.5 Film Was Made

This clip is fifteen seconds of a woman rescuing a helpless version of herself from a sea cliff, and it was made from one frontal portrait and one prompt. Seedance 2.5, 9:16, 1080p, audio on, no input video, 4,950 credits. The prompt is written as six timed shots, and the model held the same face across two characters for the full fifteen seconds.

What Was Set

Everything else is in the prompt.

Why One Photo Is Enough

The reference is a single frontal portrait: face forward, plain background, freckles visible.

Seedance 2.5 reads identity from a reference and carries it into every character that points back to it. The prompt says so directly, use the attached model face for both women, and both the rescuer in the ivory gown and the soaked figure on the cliff come out as the same person. The freckles survive the tears, the wet hair and the wide shot.

This is the whole reason the clip works as a story. Two actors would have made it a rescue. One face makes it a woman saving herself.

How the Prompt Is Built

The prompt is long, and its length is structure rather than padding. It has four parts.

Identity, once. One line telling the model to use the reference for both characters and to keep features, skin tone and age identical.

Each character, in full, before any shot. The rescuer gets a paragraph: the gown's pleats and neckline, the feather wings, the hair, the expression. The helpless self gets another: the coarse olive blouse, the worn belt, the soaked skirt, the tangled hair, the fear. The shots that follow refer to them by role and never redescribe them.

Six shots with second ranges. 0–3 hanging, 3–5 cut to the face, 5–8 reverse angle to the rescuer, 8–10 hands, 10–13 the pull, 13–15 the embrace. Each block names the framing, the camera behaviour and the action. The 3–5 block is the most specific: cut directly from the side view to a tight, nearly frontal close-up… without a zoom, swooping camera move, orbit, or transition effect. That is a hard cut, asked for by name, and it arrived at three seconds.

Realism and an avoid list. Pores, tears, hands, wet stone, fabric movement, overcast light, and only diegetic sound. Then the things that must not happen: identity changes, hidden eyes, flying, glow, speech, text.

Shot by Shot

0–3 seconds. Medium-wide side view, both hands on the ledge, feet searching. The camera holds. She lifts her head toward the person above: the movement that motivates the cut.

3–5 seconds. The cut lands as written: frontal, slightly above eye level, ocean blurred behind. Tears, lifted brows, trembling lip. No zoom, no move.

5–8 seconds. Low angle from the hanging woman's eyeline. The same face, now composed, in the ivory gown, reaching down with a small proud smile.

8–10 seconds. Back to the tearful face as fear turns to recognition. The prompt asked for a separate close-up of the hand first; the model folded that beat into this one.

10–13 seconds. Medium side shot. The rescuer braces and pulls; the other woman gets a forearm and a knee onto the ledge. Weight and effort read as real.

13–15 seconds. A full-body wide from far back: both figures embracing on the black rock against the overcast sky, pulling slowly farther away. The clip ends there, as instructed.

Step-by-Step: Make One Yourself

Step 1: Pick one frontal reference

A clear portrait, face forward, even light, plain background. One is enough, and one is better than three, several references invite the model to average them.

Step 2: Describe each character once

Before any shot, a block per character: what they wear, in fabric-level detail, their hair, their expression, their posture. Write it as if briefing a costume department.

Step 3: Write the shots as second ranges

Divide the duration into blocks and give each a range. For each block: the framing, what the camera does, what the subject does, and how it ends. Name the cuts you want, cut directly, and name the moves you do not.

Step 4: Add the realism block and the avoid list

Skin, hands, fabric, light, sound. Then a plain list of failure modes to refuse. The avoid list in this prompt is nineteen items long, and every one of them is a thing video models actually do when not told otherwise.

Step 5: Set it and render once

9:16, 15 seconds, 1080p, audio on, reference attached, input video empty. Draft at 480p if you are unsure about the timing: it is less than a fifth of the price, then render the final at 1080p.

The Full Prompt

Create an ultra-photorealistic cinematic video, exactly 15 seconds, vertical 9:16, of a Korean woman rescuing a helpless version of herself from a stormy coastal cliff. Tell the story through facial expressions, eye contact, and physical movement. No dialogue, spoken words, voiceover, or lip-sync.

Identity: Use the attached model face for BOTH women. Maintain identical facial features, skin tone, and age throughout. They are two versions of the same person.

Confident rescuer: She wears an elaborate ivory-white couture gown. The fitted bodice has fine, sculptural pleats curving upward over the bust, a deep sweetheart neckline, and fabric gathered diagonally across the waist. A long, softly draped skirt features a high thigh slit. Large, wing-shaped structures attach to the upper back and extend outward and upward behind both shoulders. Dense short white feathers cover their bases, while long, curved feathers overlap along the edges and sweep downward behind her. These are physical costume pieces that stir gently in the wind. Her black hair hangs loose in soft natural waves, keeping her face visible. She looks compassionate and confident, with a small, proud, reassuring smile.

Helpless self: She wears a weathered olive-brown blouse made from coarse, crinkled fabric, with irregular cap sleeves, a frayed V-neckline, and overlapping front panels tied into a rough knot with a hanging fabric strip. A wide, worn dark-brown leather belt with a large rectangular aged-metal buckle cinches her waist. Her long, faded brown skirt has heavy textured fabric, uneven gathers, and a worn hem. Her clothing is soaked and dirty. Her loose black hair is wet and tangled, with strands against her cheeks but neither eye obscured. She looks exhausted, terrified, and deeply helpless.

0–3 seconds — Hanging shot and head movement: Begin with a steady medium-wide side view showing her hanging below the wet black cliff ledge, both hands gripping the edge, arms bearing her weight, body suspended, and feet searching for support. Include the steep drop and rough ocean below. She initially bows her head close to the rock, struggling. Then she slowly lifts her head and turns her eyes upward toward the person above. Keep the camera largely stationary, with only subtle natural drift, so her head movement motivates the next shot.

3–5 seconds — Direct cut to her face: As she finishes lifting her chin, cut directly from the side view to a tight, nearly frontal close-up from slightly above her eye level. Match her head position and upward gaze across the cut. Her face fills most of the frame, with the ocean softly blurred behind her. Hold the camera steady. Tears trail down her cheeks, her eyebrows draw upward, and her lower lip trembles as she silently pleads with her eyes. Use a clean cinematic cut, without a zoom, swooping camera move, orbit, or transition effect.

5–8 seconds — Reverse angle to the rescuer: Cut to a low-angle close-up from the hanging woman's perspective. Her identical self crouches above in the ivory gown, white feather wings framing her shoulders. She looks downward toward the same eyeline. Hold on her face as her expression softens into a gentle, quietly proud smile, with tender eyes and subtly lifted cheeks. She reaches down without speaking.

8–10 seconds — Hands and recognition: Brief close-up of the offered hand. The helpless woman releases one hand only when help is within reach, keeping her other hand on the rock. Their hands clasp firmly. Cut back to her tearful face as recognition and fragile hope soften her fear.

10–13 seconds — Physical rescue: Cut to a medium side shot. The rescuer braces her feet and pulls with visible effort. The helpless woman pushes her free forearm onto the ledge, brings one knee over the edge, and climbs up. Show believable weight, muscle tension, balance, and continuous contact. Keep the gown clear of their footing. They move safely back from the edge.

13–15 seconds — Distant hug: Cut to a distant, full-body wide shot showing both women embracing tightly on the cliff. The rescued woman buries her face against her stronger self's shoulder, and the rescuer wraps both arms around her protectively. Show their complete figures, the ivory gown and feather wings, and the worn brown outfit against rugged black cliffs, crashing waves, and an expansive overcast sky. Slowly pull farther back as they hold the embrace. End on this distant hugging shot.

Realism: Natural skin pores, realistic tears, subtle expressions, anatomically correct hands, detailed wet stone, and believable fabric and feather movement. Soft overcast daylight keeps faces clearly readable. Muted ocean colors and natural motion blur. Sound only of waves, wind, breathing, and clothing movement.

Avoid: changing identities, hidden eyes, exaggerated crying, smug smiles, plastic skin, malformed hands, effortless lifting, teleportation, wing flapping, flying, magical glow, clothing changes, speech, music, text, or watermarks.

What It Costs

Seedance 2.5 is priced per second of output, and the rate changes with resolution. Fifteen seconds at each tier:

The jump to 1080p is not linear: it is roughly two and a half times 720p, because the model is generating four times the pixels. Which is the argument for drafting cheap: a 480p pass checks the timing and the cuts for less than a fifth of the final price, and the prompt does not change between tiers.

Where It Fell Short

Two things did not arrive as written, and both are worth knowing before you copy the prompt.

The wings. The prompt asks for large, wing-shaped structures that extend outward and upward behind both shoulders. The result is feathered shoulder pieces, dense, white, and clearly part of the costume, but a cape rather than wings. The model softened the most fantastical element into something it could keep physically plausible across the pull and the embrace. If the wings matter, expect to fight for them, or accept the version the model will actually hold for fifteen seconds.

The hand close-up. The 8–10 block asks for a brief close-up of the offered hand, then a cut back to the face. The model went straight to the face. The recognition beat is there; the insert is not. Short inserts inside a two-second window are the thing this model most often drops.

One more, about the prompt itself: the version that was actually sent began with a note, the reference holds a steady side view as she raises her head… I've built that transition into the full prompt: that was written for a person, not the model. It did no visible harm, but a prompt should contain only what the model is meant to render. The version above has it removed.

Where You Still Regenerate

A timed prompt gets the structure right the first time. It does not make the render deterministic. Run this prompt again and the face will hold, the cuts will land at the same seconds, and the cliff will be a different cliff. If a specific frame matters: a particular expression, a particular composition: that is a second render, not an edit.

Try It

Press Recreate on the prompt above. Swap the reference for your own frontal portrait, keep everything else, and start at 480p.