Seedance 2.5 Field Guide

Every test here stars Mr. Snuggles. Mr. Snuggles' voice: my younger brother, Jon Daws, who's been doing this cat for 30 years. I sampled it into ElevenLabs.

If you read nothing else

3 things to know

Pictures hold a character.

Give one clean picture per face, name it (@Image1 is MAYA) and say what it's for. Words set the scene; pictures keep the face.

Draft at 480p. Finish only the keeper.

Every run is a fresh take, so draft 2–3 and finish the winner.

Change the shot size at every cut.

Wide, then medium, then close: a classic editing rule that keeps every cut fresh. And time your lines.

The basics

Every shot, by the numbers

Length4–30 sPer generation, at every resolution. Or set the length to -1 (meaning "you choose") and the model picks it.

Resolution480 · 720 · 1080720p is the default. Native 1080p comes out 10-bit H.265 (HEVC). For playback everywhere, Chrome and Firefox on Windows included, Segmind suggests asking for H.264.

4K4K finalBytePlus calls Seedance 2.5 4K “the most capable model in the Seedance 2.5 generation”: draft at 480p, then a 10-bit 4K final (BytePlus blog, Oct 27, 2026). Segmind's Upscaled version is different: it renders at 1080p, then upscales (8-bit H.264).

Frame rate24 fpsFixed at 24, the film standard.

SoundBuilt inDialogue with lip-sync, effects, ambience and music, made with the picture.

CutsMulti-shotWrite "Shot 1:, Shot 2:" in one prompt and it cuts inside one generation.

16:9

9:16

1:1

4:3

3:4

21:9

adaptive

Also worth knowing

Task type. On fal and BytePlus you can set it: auto, reference, edit or extend. (fal, BytePlus docs)
Chain shots. return_last_frame on Segmind and BytePlus hands back the last frame, to start the next shot from. (Segmind, BytePlus docs)
Output. H.264 or H.265. .mov output (for edit and extend round-trips) on BytePlus and Segmind. High-bitrate output on fal and Segmind.
Download right away. Finished links expire: 24 h on BytePlus, 1 h on Segmind.
Lip-sync. BytePlus lists 11 languages: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean. (BytePlus docs)
Steer depth, pose, camera and lip-sync with the prompt and your references. It all lives in your words and pictures.
Seedance 2.5 flash is “now available” on BytePlus, with online use and API calls from Nov 15, 2026. (BytePlus docs)
Seeds. Segmind's seed defaults to a fixed 42 unless you change it. fal notes the same seed gives a fresh take each time. (Segmind)
Moderation. On Segmind, skip_moderation is on by default; baseline safety rules still apply. (Segmind)

Adaptive matches whatever you feed it. It's the only ratio allowed when you start from a first frame, or when you edit or extend a clip.

The options

Six ways to use Seedance 2.5

Six main ways to use it, numbered 1 to 6. Under each are its variations: other things you can feed the same mode. Each one shows what you give it, what you can make with it, and a tip to get you going. Tap one to open it.

Make a shot

Text to video

Type a scene and get a shot with sound. The fastest way to try an idea.

You give

“Your
words”

Prompt

You get

Your blocking and every line of dialogue, from words alone.

Tip: Words set the scene. To keep the exact same character in every shot, add pictures (3, References).

How it works

Good to know: each take imagines the character fresh. Add pictures to keep the same one across shots.

What you can send with it

Prompt only
Anything else turns it into a References shot

See the takes

Both ran on Segmind, one take each. Words only got the blocking and every line right, but a different cat, moon and skyline. Words can’t hold a character.

Example prompt

Night, a city street after rain, 1990s 35mm film look. Medium shot on a 50mm lens, slow push in.
A WOMAN in a yellow raincoat waits alone at a bus stop. She checks her watch, looks up the empty street, checks her watch again.
Only these sounds: steady rain, one distant car passing. No music. No subtitles.

Creator tips (4)

Put subtext under the dialogue and time each beat; the emotion lands faster. — Kyle Salazar
Keep reminding it faces stay "continuously alive and expressive". — Heavy Pulp
Direct with small physical details: a finger, a breath, a blink. — gen.ericai
Repeat what matters. Names, props and places in CAPS. — Jordan Chesney

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

First + last frame

Give it a first frame (and a last) and it makes the shot in between.

You give

“Your
words”

Prompt

His voice, second pass

You get

It opens on the empty desk, he leaps in, says the bit in his own voice and ends paws up.

Tip: For a talking shot, run it silent, then add the voice in a second pass. For a smooth push-in, make the last frame a crop of the first.

How it works

Use it for: nailing an exact composition, or bridging two shots you already have. When the character matters more than the framing, use References instead.

A last frame pairs with a first frame.
On Segmind, a first frame goes on its own, without reference images.
The ratio follows your frame (adaptive).
On Dreamina, true first + last mode takes just the two pictures; add the voice in a second pass.
It lands both frames and imagines the journey between them. Describe the path simply.

Good to know: the first frame sets the look, so use a finished-looking picture.

What you can send with it

First frame needed
Last frame optional
Images no
Extra clips not stated
Audio not stated; none on Dreamina
Last frame back Segmind and BytePlus can return the last frame, to start the next shot from it

See the takes

With two different stills (Segmind) it starts and lands on them, but the props and skyline morph in between (6/10). Make the last frame a tighter crop of the same picture (Dreamina) and it's a clean push-in, nothing morphs (8/10). The two ran on different platforms, one take each.

Frames that really differ: he arrives

Empty desk to seated — What we got (6/10)

Both frames made from the same picture, one with him removed. It starts and ends exactly on them, camera locked, props steady. The middle is invented: he climbs up from the front of the desk and the chair morphs as he gets on (6/10).

With him talking

Silent first + last, then a voice pass — Silent pass, then a voice pass (8/10)

Both frames as references + his voice (6.5/10)

First + last, lines in the prompt (5.5/10)

On Dreamina, true first + last mode takes only the two pictures, so there is no slot for his voice. Lines written in the prompt got an invented voice an octave low (5.5/10). Sending both frames as ordinary references with his voice kept the voice, but the desk sat empty for 2 s and then he popped up (6.5/10). Best in our tests (one take each, all on Dreamina): a silent first + last pass, then a second pass that adds his real voice. He leaps in, says the bit in his own voice and ends paws up (8/10).

Example prompt

The shot opens exactly on the first frame. The camera pushes in slowly toward the front door as the porch light flickers once. The shot ends exactly on the last frame. One continuous take, no cuts.
Only these sounds: crickets, the hum of the porch light. No music.

Creator tips (3)

References give the richest quality; save start and end frames for a seamless continuation. — Kyle Salazar
2.5 keeps the input image's soft color instead of over-sharpening it. — Aze Alter
Grain in the start image steers the look. Add grain back in post. — Jordan Chesney

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References

Give it pictures, clips and voices, and your character, set and voice stay the same in every shot.

You give

“Your
words”

Prompt

Voice read

You get

Same prompt as Text to video, now with his face, his set and his voice: the biggest single jump in our tests.

Tip: Send references that match what you want now. Whatever is in the picture comes into the shot.

How it works

Use it for: keeping a character, set and voice the same across shots.

Up to 30 images, 10 clips and 10 audio files.
Call them @Image1, @Video1, @Audio1 and say what each one is for.
A character sheet holds the face and the costume. Clips can steer motion, camera or performance. A clay blockout or a depth map can steer layout. Each of these is a variation below.
There are no separate settings for depth, pose, camera or lip-sync: you steer them with the prompt and the references only.

What you can send with it

Images up to 30
Clips up to 10
Audio up to 10. Audio-only references work (BytePlus, Segmind)
First frame no

See the takes

Same prompt both times, both on Segmind. With the host picture, set picture and voice, the look holds. With words only, it drifts.

Example prompt

@Image1 is MAYA: her face, hair and green work jacket.
@Image2 is the diner where the scene happens.
@Video1 sets the camera move only; its people and colors never appear.
@Audio1 is MAYA's voice.
Medium shot. MAYA slides into the booth, sets her keys down and says: "You're late. Again."
Only these sounds: diner chatter, a coffee machine. No music.

Creator tips (4)

Headshots with a neutral face, teeth showing, so performance isn't locked and teeth stay the same. — Jordan Chesney
One face per reference, even when it's the same face twice. — Jordan Chesney
A background plate plus character refs gives the best consistency. — Kyle Salazar
Make the character on a plain white background as its own clean reference. — Dr. Dreams

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Variations

References ›

Storyboard

Advanced

Draw a storyboard and it plays the panels in order, as shots.

You give

The clean-line storyboard — Clean-line board

Voice read

“Your
words”

Prompt

You get

The clean-line board on the left (on top on a phone), the take on the right: he renders as a real felt puppet, with all six shots in order.

Or just the take

Tip: Draw your storyboard clean and it renders as a real puppet. Draw only what you want on the set, and time each shot in your words too.

How it works

It's a reference image, not a special mode, so it goes in the same request as everything else. Send the board, your character sheets, a set picture and voice clips together. Panels get read in order as shots. "15 panels" is a habit people share, not a documented limit. Fewer, clearer panels read better.

Good to know: say panel borders, numbers and sketch lines never appear. Draw in clean lines with no pencil shading. A board goes with your references rather than a first frame.

What you can send with it

Board counts as 1 of 30 images
Character sheets + set the other 29
Clips up to 10
Voices up to 10
First frame no

See the takes

In my tests (Dreamina, one take per board): pencil board 7.5/10, clean-line board 8/10. Both played all six panels in order.

The lesson: a pencil board printed its hatching on him. A clean-line board fixed it.Pencil board · 7.5/10

Pencil board take — What we got: hatching on him

Clean-line board · 8/10

Clean-line board take — What we got: real felt, no hatching

Same six panels, same references, one take each on Dreamina. With the clean-line board: real felt fur, no hatching, no outlines, all six shots in order. The cuts still landed 0.4 to 1.2 s late, so the voice runs ahead of the picture. Whatever you draw on the board ends up on the set: the band riser in panel 1 showed up in the wide.

The pencil board, panel by panel: one storyboard picture, plus his picture, the set and a voice read A six-panel pencil storyboard of Mr. Snuggles at his desk

A six-panel pencil storyboard of Mr. Snuggles at his desk

You get: six shots, in this order

Play the take, or the board and take side by side

One board, one 10 s draft, on Dreamina. All six panels in order, each at its framing, but the pencil hatching leaked into the render (7.5/10). The cuts still ran late, about a second by the end, so the last shot was cut short. A board holds order and framing, not timing: time each shot in the words too.

Example prompt: board + sheets + voices

@Image1 is the storyboard: 6 panels, read left to right, top to bottom. Each panel is one shot, in order. Match each panel's framing and camera angle. The borders, numbers and sketch style never appear; the film is photoreal.
@Image2 is MAYA's character sheet. @Image3 is LEO's character sheet.
@Image4 is the diner.
@Audio1 is MAYA's voice. @Audio2 is LEO's voice.
Panel 3: MAYA says, "You're late. Again." Panel 5: LEO says, "Traffic."
Only these sounds: diner chatter, rain on the window. No music.

Creator tips (2)

Keep boards simple: the detail is mostly for client approval. — Kyle Salazar
Make a 3×3 grid of one location, then split it into separate plates. — David Israel

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

Multi-shot

Write Shot 1, Shot 2, Shot 3 in one prompt and it cuts a whole scene inside one generation. Give it a still per shot, or describe the shots in words.

You give

Shot 1: wide
Shot 2: medium
Shot 3: close

Prompt

Voice read

You get

Every shot matches its still, with clean cuts at 3.6 s and 6.3 s.

Tip: Give it a still for each shot and every shot lands just like the picture.

How it works

Use it for: covering a whole scene in one 30 s take, with geography and acting that hold across the cuts.

Good to know: show it the wide you want, so the room stays yours. A classic editing rule: change the size at every cut, so each cut feels fresh.

What you can send with it

Words only works
Images up to 30
Clips up to 10
Audio up to 10

See the takes

Play the take

Take #1000196, words only, on Segmind. The cuts landed within a quarter second of where they were written, at the right sizes. In the wide shot it invented an audience and changed the set.

Rerun: a still per cut

With words only (Segmind), the cuts landed but it invented a human audience (7/10). Given a still for each shot (Dreamina), every shot matched its still, no stray audience (8/10). The two ran on different platforms, one take each. The close-up held two seconds of silence at the end.

Example prompt

Shot 1 (wide): the empty gym, a BOY alone at center court with a ball.
Shot 2 (medium): he bounces the ball twice, eyes on the hoop.
Shot 3 (close-up): his face as he shoots; he doesn't watch it land.
Only these sounds: the ball on hardwood, a buzzing light. No music.

Creator tips (4)

Think in 30 s chunks. Cut at safe points, never mid-line or mid-action. — Kyle Salazar
Full 30 s one-take prompt: characters interrupt each other, every line in {braces}. — Jordan Chesney
Up to 15 shots in one 30 s clip, pushing continuous movement. — James D Phillips
Whip pans made inside the model, no editing. — Christopher Fryant

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

Clay (Blender)

Build a quick grey blockout in Blender, render any angle you like, and it frames a fresh shot to match.

You give

“Your
words”

Prompt

Grey clay layout — Clay blockout (Blender)

Voice read

You get

A brand-new side angle, held for the whole 10 s, in full colour.

See it side by side: what we gave it, then what we got

Tip: Send the clay as a reference, plus a set picture to keep your props. For wild angles, try depth maps too.

How it works

A clay render is a quick blockout built in Blender: rough shapes for the set, the props and the character, placed where they go, with a camera at the angle you want. It's called a clay render because everything is one plain matte grey, like a sculptor's clay model, so all you judge is shape, layout and camera.

ByteDance also documents clay renders as references for light direction, colour, shadows and poses. (ByteDance Seed)

Block the set in grey shapes in Blender and render the exact angle you want.
Send that still as a reference for a fresh shot, with the face picture, the set picture and a voice read.
A rendered clip of moving clay can carry a camera arc the same way.

Good to know: clay shines as a reference. Used as a first frame, the grey carries into the shot.

What you can send with it

Clay still or clip as a reference
Images up to 30
Clips up to 10
Audio up to 10

See the takes

Fresh shot: framed from a clay layout

All on Dreamina, one take each. Instead of re-shooting, we built a quick grey blockout of the set in Blender and made a fresh shot. The clay still held the new side angle for the whole 10 s, with no grey leaking in (7/10). The painted first frame opened on the exact angle, but he turned to camera within a second (6/10). The moving clay clip copied the camera arc, but a prop from the host picture wandered in (5/10).

Three clay angles, one per cut: top, worm's-eye, behind

Three clay angles — What we got (6.5/10)

Three clay angles, side by side — Side by side

A multi-shot with one clay render per cut. The top-down shot matched. It mirrored the worm's-eye shot and pointed the behind shot at the backdrop (6.5/10). The lines came out the cleanest of the multi-angle tests. Three depth maps of the same angles got all three directions right (7.5/10): see Depth maps. In this one-take comparison, depth held framing better than clay.

Side by side: the clay, then the take

The reference on the left (top on a phone) switches at each cut, so you can check every shot against what we gave it.

Example prompt

@Image1 is a grey clay render of the diner (a quick Blender blockout): match its camera angle and where everything sits. The grey clay never appears; the film is photoreal.
@Image2 is MAYA: her face, hair and green work jacket.
@Image3 is the diner.
@Audio1 is MAYA's voice.
Medium shot from the side, framed as in @Image1. MAYA slides into the booth, sets her keys down and says: "You're late. Again."
Only these sounds: diner chatter, a coffee machine. No music.

Creator tips (2)

A clay render of the scene works as a layout reference. — alexgnewmedia
Turn real film shots into previz to show the model, like you would a DP. — Josh Daws

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

Depth maps

A grayscale picture, white near and black far. Give it one and a fresh shot takes that camera angle.

You give

“Your
words”

Prompt

Voice read

You get

Three wild angles, top-down, worm's-eye and from behind, one depth map per cut. All three face the right way.

Tip: Use depth maps on fresh shots, one per cut, for angles a camera rarely gets.

How it works

A depth map is a grayscale picture where white is near and black is far. Render one in Blender, or pull one out of any video with a depth model. We used Depth Anything V2 on a free stock clip. It is a common technique.

Send it as a reference for a fresh shot, with the face, the set and the voice.
Say it sets only the camera angle and where things sit; the grey never appears.
In a multi-shot, give one depth map per cut.
There is no depth setting anywhere: it's steered by the prompt and the reference only.

Good to know: a depth map carries shape, not colour or detail, so add a set picture for your props.

What you can send with it

Depth maps as images, up to 30
Clips up to 10 (a depth video works as a clip)
Audio up to 10
First frame no

See the takes

One depth map, fresh shot

One depth map, side by side — Side by side

The fresh shots ran on Dreamina, one take each. One depth map framed a fresh shot from the side. The framing matched and nothing went grey, but the desk became a round table (7/10).

Three depth maps, one per cut: front, side, close

Three depth maps: front, side, close — What we got (8/10)

A multi-shot with one depth map per cut. The cuts landed at 3.7 s and 6.3 s and every angle matched its map (8/10). In the side shot the desk turned into a table again.

Three wild angles: top-down, worm's-eye, from behind

Three wild depth angles — What we got (7.5/10)

Three depth maps, side by side — Side by side

Angles a camera rarely gets. All three went the right direction. The second cut came about 1 s late, and the shot from behind dropped the desk (7.5/10). The same three angles as clay renders scored 6.5/10: see Clay (Blender).

Depth vs clay, same three angles

Three clay renders — Clay renders (6.5/10)

Clay mirrored the worm's-eye shot and pointed the behind shot at the backdrop. Depth got all three right. One take each, so this is a lead, not a rule.

The honest contrast: a depth map on a re-shoot

Re-shoot with a depth map — Re-shoot with it (3/10)

Sent with the source take to re-shoot it from a new angle, the depth map was ignored and the framing stayed identical (3/10). That re-shoot ran on Segmind and the fresh shots on Dreamina, one take each: on our re-shoot the source video won, and on fresh shots depth maps worked.

Example prompt

@Image1 is a depth map of the shot: white is near, black is far. Match its camera angle and where everything sits. It is not a picture of the scene; nothing grey appears. The film is photoreal.
@Image2 is MAYA: her face, hair and green work jacket.
@Image3 is the diner.
@Audio1 is MAYA's voice.
Low angle from the side, framed as in @Image1. MAYA slides into the booth, sets her keys down and says: "You're late. Again."
Only these sounds: diner chatter, a coffee machine. No music.

Creator tips (1)

A depth-map video plus image refs for fights and dance. — Sam Wasserman

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

Camera move

Borrow a camera move from any clip, even an old film. Your shot copies the move.

Try it with an old film: I sent a silent black-and-white clip of a woman running at a handheld camera, from a public-domain 1968 film. He took the runner's place: he runs straight at the camera as it backs away, bouncing like the clip, in full colour.

You give

“Be the runner”

Prompt

Film clip: a woman running at the camera, sound off — Film clip, sound off (public domain, 1968)

You get

The film clip on the left (on top on a phone), our take on the right: the same handheld move, with him in the runner's place, in full colour. Only the move came through.

Or just the take

Tip: Strip the clip's sound first, so only the move comes through.

How it works

Send the clip as @Video1 and say it sets the camera move only.
Its people, colors and sound must never appear: say so, and take the sound off the file.
It goes in the same request as your face, set and voice references.
A clip with a person in it works too: say your character takes that person's place, and that the person, the place and the colours never appear.
There is no camera setting anywhere: the move is steered by the prompt and the clip only.

What you can send with it

Clips up to 10
Images up to 30
Audio up to 10
First frame no

See the takes

A plain camera clip: a slow push-in

You give

“Your
words”

Prompt

Voice read

You get

It copied the push-in, and only the push-in.

See it side by side: what we gave it, then what we got

A camera clip as a reference

The clip with its sound ran on Segmind, the silent rerun on Dreamina, one take each. A clip sent only for its camera move copied the push-in, but also copied the clip's dialogue: he said the wrong lines (3/10). Strip the sound from the clip first and it copies only the move (7/10).

Side by side: the clip, then the take

Camera clip, side by side — The silent camera clip, then our take

The camera clip on the left (top on a phone), our take on the right.

A film clip with a person in it · 8/10 · one take

He runs at the backing-up camera — Our take

Side by side: the clip, then the take

One take, on Dreamina, 8/10. The camera copy really works: same bounce, same sway second by second, and the runner's whole path through the frame carried over to a puppet in full colour, with nothing from the source clip leaking in. The clip: a woman running toward a handheld camera that backs away in front of her, sent silent and used only for the move and where the runner sits in the frame. He took her place and the camera backed away in front of him, bouncing like the clip, in full colour. Worth tweaking: partway through, the set drifted into a bigger stage. Add “the same set the whole shot” to your prompt.

Example prompt

@Video1 sets the camera move only: a slow push-in. Its people, colors and sound never appear.
@Image1 is MAYA: her face, hair and green work jacket.
@Image2 is the diner.
@Audio1 is MAYA's voice.
MAYA slides into the booth, sets her keys down and says: "You're late. Again."
Only these sounds: diner chatter, a coffee machine. No music.

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

Movement

Pull a depth video out of a dance clip and your character dances the same moves.

You give

“Your
words”

Prompt

Waist-up depth video of the clip — Depth video, cropped

You get

He copied the moves in order and on time. Clip: Mixkit, “Man dancing hip hop on a red background”, free license.

Tip: For a seated dance, crop the clip to the waist and roll a few takes. My seated take came from a waist-up depth video.

How it works

Method credit: Sam Wasserman, who uses a depth-map video for dance.
A depth video is a grey version of the clip: white is near, black is far. Any depth model can pull one out of any video. It keeps only the shapes of the movement: no face, clothes or colours.
It copies the whole body and its place in the frame, so a full-body dancer brings a standing performance. Crop the clip to the waist for a seated one.
Send it as @Video1 and say it sets only the movement.
There is no pose setting anywhere: it's steered by the prompt and the clip only.

Good to know: look over each take for anything extra from the clip, like the dancer's sleeves. A depth video has no face at all, so it sits easily with BytePlus's rule on real faces in references.

What you can send with it

Clips up to 10, all clips together up to 30 s
Images up to 30
Audio up to 10

See the takes

In my tests (Dreamina): the moves came in order and on time in every take (one only through about 5 s). Full body: he stood every time (3 of 3). Waist-up: seated in 1 of 3, with band instruments or the dancer's sleeves leaking into the others.

Our first test side by side: the cropped clip, then that take (6/10)

Retest, depth video cropped to the waist · 3 takes: 8, 4.5, 4

Waist-up depth video, take a (seated) — a · 8/10 · seated

Waist-up depth video, take b (band leaked in) — b · 4.5/10 · band set leaked

Waist-up depth video, take c (stood up) — c · 4/10 · stood, sleeves leaked

All on Dreamina. The moves came in order and on time in all three. Only take a kept him in his chair. Take b grew a band set: a bass, drums and a microphone. Take c stood up and took the dancer's short sleeves, watch and framing.

Retest, depth video of the full body · 3 takes: 4, 3.5, 5

Full-body depth video, take a — a · 4/10 · stood, set rebuilt

Full-body depth video, take b — b · 3.5/10 · stood, band set leaked

Full-body depth video, take c — c · 5/10 · stood

Three takes on Dreamina. He stood up every time (3 of 3). Take a: he stood the whole take, and the set got rebuilt around a standing performer: a couch, the desk pushed aside, no moon, and a band with upright bass, drums and keys. Its moves came in order and on time through about 5 s, with the camera locked. Takes b and c: moves in order and on time.

First test, one take: depth video cropped to the waist · 6/10

Waist-up depth video as the movement — What we got

He stays seated and the camera holds. The moves come in the dancer's order, about 2 s late at the start. Flaws: a drum kit and a bass appeared, and the tree and mug disappeared. Clip: Mixkit, “Man dancing hip hop on a red background”, free license.

First test, one take: depth video of the full body · 4/10

Depth video pulled from the clip — The depth video

Full-body depth video as the movement — What we got

Full-body depth video, side by side — Clip and take together

The moves match almost beat for beat, but he stood up out of his chair and the framing drifted with the dancer.

First test, one take: the clip itself, sound off · 5/10

Dance clip as the movement — What we got

Dance clip, side by side — Clip and take together

He stayed seated, but his moves only loosely match the dancer.

Example prompt

@Video1 is a grayscale depth video of a dancer from the waist up, used ONLY for the movement of the arms, shoulders and head; no grey ever appears, and its framing is not copied.
@Image1 is MAYA: her face, hair and green work jacket.
@Image2 is the diner.
MAYA stays seated in the booth the whole time and, from the waist up, copies the moves in @Video1 beat for beat. Locked-off medium shot.
Only these sounds: diner chatter. No music. No subtitles.

Creator tips (1)

A depth-map video plus image refs for fights and dance. — Sam Wasserman

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

A look

Borrow a look from a picture: its light and its filters, on your own set.

You give

“Worn 80s VHS”

The look, in words

The 80s VHS look picture — The look picture

Voice read

You get

One reference picture, ‘for lighting and filters only’, plus the look in words brought the whole 80s VHS feel onto his own set. Name every object you don’t want, because the mic still snuck onto the desk.

See it side by side: the look picture, then the take

Tip: Say the picture is “for lighting and filters only”, then write the look in words too, so the picture and the words pull the same way.

How it works

Make a picture that has the look you want: the light, the colour, the texture.
Send it next to your character and your set, and say it is for lighting and filters only.
List what in that picture you don’t want: its room, its furniture, its props.
Write the look as its own line of the prompt, in plain words.

What you can send with it

Images up to 30
Clips up to 10
Audio up to 10

See the takes

This take is one try on Dreamina, 7/10. Before it, I tried the look in words alone (black and white noir), and a picture of his own set graded black and white. Those takes are here too.

3 takes a side (Dreamina). Picture: 6.5, 6, 6. Words: 6, 6.5, 5. Picture: blind bars on the set in 3 of 3, never on him, and he stayed in colour. Words: black and white in 2 of 3, no blinds in any. The moon went missing in all 6 takes.

Our first test side by side: the graded picture, then that take (6/10)

The 80s VHS look: a picture for lighting and filters only, plus words · 1 take: 7/10

One take, on Dreamina. The whole VHS feel came through: soft analog video, warm tungsten glow, bloom, scan lines and tape noise, on his own set, moon and skyline. The miss: the picture’s microphone landed on his desk, even though the prompt said not to use it.

Retest, a graded picture of his own set · 3 takes: 6.5, 6, 6

Look from a graded picture of his set, take a — a · 6.5/10

Look from a graded picture of his set, take b — b · 6/10

Look from a graded picture of his set, take c — c · 6/10

All on Dreamina. Hard blind bars on the set in all three, never on him. He stayed in colour every time. The picture's couch leaked in all three, and the moon was gone.

Retest, the look in words · 3 takes: 6, 6.5, 5

All on Dreamina. Black and white in 2 of 3, but no blind bars in any, and the moon was gone in all three.

First test, one take: a graded picture of his own set · 6/10

His set, graded black and white — The graded picture

Look from his own set, graded — What we got

One take, on Dreamina. Black and white with hard blind stripes on his own skyline, moon and desk. Flaws: he stays in colour and the stripes miss him, his suit stayed blue, his tie stayed red, and the desk changed.

First test, one take: the look in words, the light described exactly · 3/10

Look in words, light described — What we got

No blind stripes at all, the tie and chair kept their colour, and the moon and cat-ear skyline were lost. More exact words made it worse.

First test, one take: the look in words · 5/10

Same: the colour drains, the light doesn't change.

First test, one take: a noir picture of another room · 3/10

Real noir light, but the picture's blinds, window and desk replaced his set. Telling it “nothing in that picture appears” did not stop that room's window and desk from coming in.

First test, one take: a soft grainy-film picture · 4/10

Multi-shot with a look picture — What we got

The look barely changed.

The prompt for this take

@Image1 is SNUGGLES, a grumpy grey tabby stop-motion puppet in a navy suit and loose red tie, the only character: use it for his appearance only. @Image2 is his set: walnut desk, teal chair, cat-ear city skyline, crescent moon, cat mug. Use it for the layout of the place only. Refer to @Image3 for lighting and filters only. Do not use its room, its desk, its armchair, its plant, its lamps or its microphone.
LOOK: the whole shot looks like a frame off a worn 1980s VHS tape of a late-night TV broadcast, shot on a tube video camera: soft, slightly blurry analog video, warm orange-magenta tungsten light, glowing blown-out highlights with soft bloom, murky low-contrast shadows, colour bleeding sideways past edges, faint horizontal scan lines, light tape noise, and a thin wobbly tracking line near the bottom of frame.
A locked-off medium shot of SNUGGLES behind his desk.
@Audio1 is SNUGGLES's voice saying these lines, in order: "The dog gets a walk." "The dog gets a treat for SITTING." "Sitting!" "I sit all day!" "Where's MY parade?"
He gestures with one paw, throws both paws up on "Sitting!", jabs his chest on "I sit all day!", narrows his eyes on "Where's MY parade?" and holds the stare. His mouth moves with every syllable.
[Strictly exclude] armchairs, potted plants, wall lamps, microphones, people, sharp modern digital video, crisp HD detail, on-screen text, logos.
Only these sounds: his voice, quiet room tone, a studio audience laughing at the end. No music. No subtitles.

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

Character sheet

Two character sheets: the head on one, the body with no head on the other. One face to copy, so he stays himself in every shot.

You give

“Your
words”

Prompt

Voice read

You get

A strong likeness, every line in order, and his head sits right on the body.

Tip: Send a set picture with the sheets for a full, rich set.

How it works

How to make the sheets

Plain grey background. Nothing else in the picture.
A row of turnarounds: front, three-quarter, side, back.
The head on its own sheet. The body, without the head, on another.
One face per reference. A second face gives it two to copy.

How to send them

Name both: @Image1 is his head, @Image2 is his body, and his head is always the one in @Image1.
Add the set picture and the voice in the same request.

What you can send with it

Sheets count as 2 of 30 images
Clips up to 10
Audio up to 10

See the takes

In my tests (Dreamina, one take): 7.5/10.

A head sheet + a headless body sheet

One take on Dreamina, with the two sheets, the set picture and his voice read. Strong likeness, every line in order, and no mix-up between head and body (7.5/10). The set came out thinner than the set picture.

The prompt we sent

RULE 1: SNUGGLES's head, face and fur are always exactly as in @Image1. RULE 2: his suit, shirt, tie, paws and tail are always exactly as in @Image2; the head in every shot is his head from @Image1, joined naturally to that body. RULE 3: the set is always exactly as in @Image3. @Audio1 is SNUGGLES's voice saying these lines, in order: "The dog gets a walk." "The dog gets a treat for SITTING." "Sitting!" "I sit all day!" "Where's MY parade?" SNUGGLES sits behind the walnut desk and rants: he gestures with one paw, throws both paws up on "Sitting!", jabs a paw at his chest on "I sit all day!", and narrows his eyes on "Where's MY parade?", then holds the stare. His mouth moves with every syllable. Medium shot, locked-off desk camera, warm late-night key light. Only these sounds: his voice, quiet room tone, a studio audience laughing at the end. No music. No subtitles. No on-screen text.

Example prompt

@Image1 is MAYA's head sheet: her face and hair from several angles.
@Image2 is MAYA's body sheet: her green work jacket, jeans and boots. It has no head; her head is always the one in @Image1.
@Image3 is the diner.
@Audio1 is MAYA's voice.
Medium shot. MAYA slides into the booth, sets her keys down and says: "You're late. Again."
Only these sounds: diner chatter, a coffee machine. No music.

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

References ›

A still from a take

Love an earlier take? Pull one still from it and send that.

You give

“Your
words”

Prompt

One still from
your take

The still

Voice read

You get

The right words, a clean set and livelier acting.

Tip: Send the still with the set picture and the voice read, and write the new lines fresh.

How it works

Pick a frame from the take that shows the character and the look you want.
Send it as an image with the set picture and the voice read.

What you can send with it

Images up to 30
Clips up to 10
Audio up to 10

See the takes

Reuse an earlier take as the reference

One still from a take — One still from the take + set + voice (8/10)

All on Dreamina, one take each. We wanted the Dog bit from an earlier Tree take. Sent as the only reference, the take with its sound copied the old lines (4/10). With the sound stripped the words were right, but the old Christmas tree leaked in and he went stiff (6.5/10). One still pulled from the take, plus the set and his voice, got the right words, no leaks and livelier acting (8/10). That still is Jordan Chesney’s method.

Tip: To reuse a take, try a still pulled from it. In our one take each, the video reference dragged its words and props along.

Example prompt

@Image1 is a still from an earlier take: MAYA's face, hair, green work jacket and the light. Only the look comes from it, not the action.
@Image2 is the diner.
@Audio1 is MAYA's voice.
Medium shot. MAYA slides into the booth, sets her keys down and says: "You're late. Again."
Only these sounds: diner chatter, a coffee machine. No music.

Creator tips (1)

Pull one still from the take and use it as the reference. — Jordan Chesney

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Change a take

Edit

Edit a shot: change one thing in words: a costume, a prop, the weather. Everything else stays.

You give

“Make his tie green”

What to change

You get

Only the tie changed. Face, timing, set and even the sound stayed the same.

Tip: Name everyone and say exactly what stays, and only that one thing changes.

How it works

Source clip 4–30 s. Output keeps its length and framing, though it can come out up to 0.4 s shorter (BytePlus docs).
Swap a costume, a prop, a background, the weather.
Turning sound off strips the audio track.
Fal and BytePlus let you set the task type (auto, reference, edit or extend) instead of leaving it to the model.
On Segmind, an edit or extend is billed as at least 50 seconds of video (source plus new) (Segmind).
Edit with a picture, a new place, a new angle and the sound each have their own variation below.

What you can send with it

Source clip 4–30 s
Images up to 30
Audio up to 10
Extra clips Segmind no; on fal, not documented (our tests only)
First frame no

See the takes

One take, on Segmind (9/10). Only the tie changed. Face, timing, set and even the sound stayed the same.

Example prompt

@Video1 is the master. Change only one thing: the MAN's jacket goes from navy to maroon. There are exactly three people. The other two people, every face, the room, the camera, the timing and the sound stay exactly as in @Video1.

Creator tips (2)

When a refusal makes no sense, cut the line that likely caused it, run the rest, then add it back with edit. — Kyle Salazar
Add "preserve the proportions" of the rest of the body. — Christopher Fryant

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Variations

Edit ›

Edit with a picture

Attach a picture of a prop, and that exact prop drops into your shot.

You give

“Swap in this mug”

What to change

You get

It became that exact mug, logo and all, through the sip. Nothing else moved.

Tip: Name the one thing that changes and say everything else stays as in the take.

How it works

Same as an edit in words, plus a picture of the new prop.
Say which prop it replaces and that nothing else changes.

What you can send with it

Source clip 4–30 s
Images up to 30, one of them the prop
Audio up to 10

See the takes

Edit with a picture

One take, on Segmind. We attached a picture of a mug and it became that exact mug, logo and all, through the sip. Nothing else moved (8/10).

Example prompt

@Video1 is the master. @Image1 is the new coffee mug. Replace only the MAN's coffee mug with the mug in @Image1, logo and all. Every face, the room, the camera, the timing and the sound stay exactly as in @Video1.

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Edit ›

A new place

Give it a picture of a diner and the whole show moves there. Same performance, new set.

You give

“Move him to this diner”

What to change

You get

The whole set swapped for the diner. Near frame for frame, same words, same timing.

See it side by side: your shot, then the same shot moved to a diner

Tip: Show it a picture of the new place, and the set matches the picture.

How it works

Use it for: keeping a take you love and moving it somewhere new. It's an edit (mode 4), so the take keeps its length, framing and sound.

Fence the place picture off in the prompt, the way you would a look picture: it “shows only what the place looks like… It is NOT a shot and NOT a camera angle; never copy its framing.”

What you can send with it

Source clip 4–30 s
Images up to 30, one of them the new place
Audio up to 10

See the takes

Set swap with a picture — With a picture of the diner (9/10)

Set swap, words only — Words only (7.5/10)

Both on Dreamina, one take each. With a picture of the diner, the old set was fully gone, the diner matched the picture, and he played it near frame for frame with the same words and timing (9/10). Our best edit so far. With words only it also worked, but it made up details we never asked for, snow on the sill and a long booth, and one gesture came early (7.5/10).

Example prompt

@Video1 is the master. @Image1 is the new location: a late-night diner booth. Move MAYA into @Image1. Her performance, every line, the timing, the camera and the framing stay exactly as in @Video1. Nothing from the old room appears.

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Edit ›

A new angle

The same performance, from a new camera angle.

You give

“Low, from the side”

The angle

You get

The camera drops low to a three-quarter view, with the acting and timing intact.

Tip: Show it a picture of the angle. For a big move, make a fresh shot from a depth map or a clay blockout (see References).

How it works

Add a character sheet or a set picture as references to hold the face and the room from angles the take never showed.

Good to know: check eyelines and continuity before you cut the angles together.

What you can send with it

Your take as @Video1
Images up to 30
Extra clips up to 9 (yours is the 10th)
Audio up to 10

See the takes

In my tests: Segmind re-shoots, one take each (words, a picture of the angle, a depth map), then one Dreamina take that asked for four new setups in words (6/10).

It went low as asked, but straight-on instead of from the side. The acting and timing held exactly.

Rerun: show it the angle

A picture of the new angle was the only thing that moved the camera: low three-quarter, not a full side view (6/10). A depth map was ignored, the framing stayed identical (3/10). On a re-shoot the source video wins. On a fresh shot, depth maps do work: see Depth maps.

Four new setups in words alone · one take, Dreamina · 6/10

Re-shot as four setups, words only — What we got: four sizes, all from the front

The four shot sizes in that take

Close-up from the words re-shoot — Close-up

Extreme close-up from the words re-shoot — Extreme close-up

What I got: words alone cut one locked take into four shot sizes over the master's audio, word for word, but the camera stays in front: “from behind” and “side profile” both came back as front views, and the cuts drifted up to 1.75 s late. The wide also invented armchairs, pillars and a curtain.

Example prompt

Re-shoot @Video1 from a new angle: low, over the driver's shoulder, looking through the windshield. Same performance, same lines, same timing. Nothing else changes.

Creator tips (3)

Add the character sheets and background images alongside the video reference. — Jordan Chesney
One solid video, the character images and a short prompt of about 15 s, then cut it in. — Kyle Salazar
Check eyelines and continuity before you cut. — Olga Baranova

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Edit ›

The sound

Advanced

Same picture, new sound. Strip a music bed, dub the lines into another language, or lip-sync him to your own voice file or a song.

You give

“Same lines, in Spanish”

What to change

Spanish read (ElevenLabs)

You get

A Spanish dub: a Spanish track from ElevenLabs, and Seedance matches his mouth to it and keeps the picture.

Remove the music: attach the clip and write: “Remove the background music from @Video1. Keep the dialogue, the picture and the timing unchanged.” That's the docs' wording, and it came back clean, with the voice, words and timing kept. Render a few and pick.

Another language: ask for the same lines in Spanish and Seedance dubs them on its own. To keep his exact voice, send a Spanish track from ElevenLabs as @Audio1, and Seedance matches his mouth to it.

Lip-sync to your own track: send your finished voice file or song as @Audio1 and say to use it exactly as it is. It kept the recording as it was, the same voice, and the mouth moved well, shutting in the pauses. It sings too: the mouth lands on the words and the band comes through.

You give

“Use @Audio1 as is”

Prompt

Voice read (a real recording)

Song (ElevenLabs Music)

You get

What I got: he sings the verse. I made the song with ElevenLabs Music, and Seedance lip-synced him to it. The mouth lands on the words and the band comes through.

And to a spoken voice file: his real recording, kept as is

How it works

Removing music by edit is documented by BytePlus; its example is “remove the background music from @Video 1” (BytePlus docs). Use that wording.
Asking for the same lines in another language can give a dub with matching lips. Seedance can speak the new language on its own; for an exact voice, send a read in the voice you want.
Dubbing comes from our own tests; the docs cover music removal.
Lip-sync to your own track is a fresh shot, not an edit: send the track as @Audio1 with his face and the set, and say to use it unchanged. Check the timing against your file.
BytePlus lists 11 lip-sync languages (BytePlus docs). There is no lip-sync setting: it's steered by the prompt and the voice reference.

Good to know: compare frames with the original before you keep it.

What you can send with it

Source clip 4–30 s
Images up to 30
Audio / voice refs up to 10

See the takes

In my tests (Dreamina): removing the music with the docs' wording came back clean in 2 of 3 takes (9, 4, 8.5); one kept the music for its first 4 s. Spanish by Seedance alone: 2 of 3 spoke Spanish, one started in English (3, 5, 7). Best dub: an ElevenLabs Spanish read in my brother's cat voice, lip-synced by Seedance (7/10, one take; the punchline came out half English, as a broken “entrapp-a”). Lip-sync to my own track, one take each: his real voice file 7/10, the song 5/10.

Lip-sync to your own track · one take each: 7 and 5

Lip-sync to his real voice file — Spoken voice file · 7/10

He sings to the song, lip-synced — Singing · 5/10

Both on Dreamina, one take each. Spoken (7/10): it kept my brother's real recording, the same voice with no re-voice, and the mouth moved well, shutting in the pauses. Worth tweaking: it re-timed the recording about 13% slower, with longer pauses, and added a laugh at the end. For tight timing, use the take's own audio in your edit, and check each word of a song against your track. Singing (5/10): the song was made with ElevenLabs Music, the vocal swapped into my brother's cat voice with speech-to-speech. The mouth lands on the words and the band comes through. Worth tweaking: it re-sang one line into the instrumental break, the framing drifted in to a close-up, and the moon went missing.

Remove a music bed, the docs' wording · 2 of 3 clean

Before: the take with a music bed under it — Before: a music bed under the take

After: music removed, take a — After, take a · 9/10 · clean

Prompt: “Remove the background music from @Video1. Keep the dialogue, the picture and the timing unchanged.”

The other two takes

Music removal, take b (music kept for 4 s) — b · 4/10 · music stayed for 4 s

Music removal, take c — c · 8.5/10 · clean

All on Dreamina. In a and c the bed was gone, and the voice, words and timing stayed. The picture was redrawn with a slight reframe. b kept the music for its first 4 s.

Spanish by Seedance alone (lines in the prompt, no Spanish audio) · 3 takes: 3, 5, 7

Spanish from the prompt alone, take c — c · 7/10 · every line, a voice close to his

Spanish from the prompt alone, take b — b · 5/10 · every line, a much deeper voice

Spanish from the prompt alone, take a — a · 3/10 · English for the first 8 s

All on Dreamina: the English take, plus the Spanish lines written into the prompt. Takes b and c said every line correctly. Take a started in English, for its first 8 s. b switched to a much deeper voice; c kept a voice close to his.

Best dub: a Spanish read made in ElevenLabs, lip-synced by Seedance

Asked for the same lines in Spanish (Segmind): correct Spanish and the lips re-synced, but an invented voice (7/10). Then ElevenLabs made a Spanish read in my brother's cat voice, and Seedance lip-synced his mouth to it and kept the picture (Dreamina, 7/10). Flaw: the punchline came out half English, as a broken “entrapp-a”. The two ran on different platforms, one take each. Seedance did not make that voice; ElevenLabs did.

First try, other wording: the music stayed (2/10)

Before: a music bed mixed under the take

One take, on Dreamina. We mixed an instrumental bed under a take and asked it to take the music out. The bass and drums were still there. It re-rendered the sound instead of separating it (2/10). The docs' wording, above, fixed it in 2 of 3 takes.

Example prompt: remove the music

Remove the background music from @Video1. Keep the dialogue, the picture and the timing unchanged.

Example prompt: lip-sync to your track

@Audio1 is the finished soundtrack of this shot. Use it exactly as it is, unchanged: the same voice, the same words, the same timing and pauses. Do not re-record or re-voice it.
@Image1 is MAYA: her face, hair and green work jacket.
@Image2 is the diner.
A locked-off medium shot of MAYA in the booth. She lip-syncs to @Audio1: her mouth opens and closes on every syllable, in time, and shuts in every pause.
No subtitles. No on-screen text.

Example prompt: a dub

Edit @Video1. Same picture, same timing, same framing. MAYA says the same lines in Spanish, in the voice of @Audio1. Her lips match the new words. Keep the room tone. No music.

Creator tips (6)

A 14 s "rainbow passage" read as the voice ref: "The voice of the character (as shown in @Image1) is @Audio1." — Jordan Chesney
Use a custom ElevenLabs voice as the ref, with the voice description at the very top of the prompt. — Kyle Salazar
With several voices, check each lands on the right character and the lips move with it. — Dr. Dreams
Always write "no music". Pull the vocals in Resolve and add your own effects. — Dr. Dreams
Mispronunciations? Re-run the lines through Seed Audio with the voice refs and swap them in. — Christopher Gwinn
On a dub, listen for the accent: one mixed European Portuguese with Spanish. — alexgnewmedia

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Extend

Extend forward: add a new 4–30 s after a clip you already have. Extend backward is the variation below.

You give

“Continue it”

What happens next

You get

The new part starts almost exactly on the old last frame. The join is hard to see.

Tip: End the source mid-action with the camera moving, then join the new part in your editor.

Add a character: paint a new character into a frame of your take and send it with the take. He appears exactly where you put him, in frame the whole shot.

You give

Placement picture: the drummer painted in at the desk — Placement picture

“Add him here”

Prompt

You get

What I got: the drummer cat showed up right where the picture put him, fully in frame and the same size as the host, and stayed there the whole shot. The host's words and timing held.

How it works

Source can be as short as 2 s. You get only the new 4–30 s, so join it to the original in your editor. Takes references too: up to 30 images, 9 extra clips and 10 audio files.

What you can send with it

Source clip 2–30 s
Images up to 30
Extra clips up to 9 (yours is the 10th)
Audio up to 10
Task type on fal and BytePlus, set it to extend

On Segmind, an extend is billed as at least 50 seconds of video (source plus new).

See the takes

Play the join (last 4 s of the original, then the extension)

One take, on Segmind. The new part starts almost exactly on the old last frame. The join is hard to see.

Extend with a new character

The drummer from the band picture walked on and matched it, but stayed half off the edge, and the host only side-eyed him (Segmind, 7/10). Writing “fully in frame” and a glare didn't fix it (Dreamina rerun, 3/10). Different platforms, one take each: in this rerun, words didn't place him. A placement picture did, below.

Add a character with a placement picture · one take

The drummer added, right where the picture put him — What I got · 5/10

One take, on Dreamina (5/10; the words-only rerun above scored 3). I sent the take and a frame of it with the drummer painted in where I wanted him. He appeared exactly there, fully in frame, and stayed the whole shot. Worth tweaking: the rimshot and the host's glare didn't land, so those need prompt tweaks.

Example prompt

Seamlessly continue @Video1 from its last frame. Same camera move, same speed. She keeps walking down the hall, stops at the last door and knocks twice.

Creator tips (4)

Open with "Seamlessly continue @Video1", then "00:01–00:XX CUT TO:" and the new shot, so it moves straight on to the new shot. — Jordan Chesney
Feed the previous 30 s back plus the character images. The video carries the sound, tone and light. — Kyle Salazar
Ask it to keep the clip as is and add seconds, for a seamless join. — Olga Baranova
15 s feels rushed. Run 1-minute long takes, then fill in with inserts. — Christopher Gwinn

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Variations

Extend ›

Extend backward

Our tests, not documented

Add an establishing lead-in before a clip you already have.

You give

“Add a lead-in”

What comes before

You get

The lead-in ends very close to the original first frame.

Tip: Roll a couple of drafts and pick the smoothest join.

How it works

Documented only on Segmind, as an "establishing lead-in", and in BytePlus's wording (Segmind, BytePlus docs). What you see here is our tests.

What you can send with it

Source clip 2–30 s
Images up to 30
Extra clips up to 9 (yours is the 10th)
Audio up to 10

See the takes

Play the join (the lead-in, then the original)

One take, on Segmind. The lead-in ends very close to the original first frame; his head tilt is a touch off.

Example prompt

Add an establishing lead-in before @Video1: a slow aerial drift down the street to the house, ending exactly where @Video1 begins.

Creator tips (1)

Same trick backward: keep the clip, add seconds before it. — Olga Baranova

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Finish

Draft, then final: make it at 480p. If you like it, finalize the same take at 1080p. Upscaling is the variation below.

You give

Finalize
at 1080p

One click

You get

The same moment up close: the 480p draft on the left (on top on a phone), the 1080p final on the right. Same performance frame for frame, now sharp.

Or the whole 1080p final

Tip: Draft 2–3 takes, then finalize the keeper.

How it works

Make it at 480p. If you like it, finalize the same take at 1080p.
The final keeps the draft's prompt, references, length, ratio, seed, audio setting and task type, and comes out at 1080p only (Segmind, fal).
How long a draft stays good: 14 days on Segmind. On fal, US-region drafts last 7 days. On BytePlus, 14 days for draft IDs made from Oct 28, 2026 (older ones, 7) (Segmind, fal, BytePlus docs).
Download the final right away: links expire after 24 h on BytePlus and 1 h on Segmind.

Good to know: finalizing the same draft twice may give you two finals (untested).

What you can send with it

Nothing new reuses the draft's prompt and every reference

See the takes

One take, on Segmind. Same performance frame for frame. The fine detail is redrawn, though: fur, teeth and eyelids change a little (7/10).

Creator tips (2)

Draft mode is "a lifesaver" for iterating. — Heavy Pulp
Always test in draft first, then finish the same prompt and seed. — Kyle Salazar

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Variations

Finish ›

Upscale

Sharpen a finished shot one step up: 720→1080, 1080→4K.

You give

Sharpen
in Topaz

Our test

You get

Same picture and sound, only sharper.

Tip: We used Topaz. Check skin, grain and fabric up close, where upscaling shows most.

How it works

Segmind's Upscaled version renders a step lower, then sharpens up: 480→720, 720→1080, 1080→4K. Its 4K renders at 1080p and is then upscaled, and the output is 8-bit H.264 (Segmind). BytePlus has a separate Seedance 2.5 4K model with a 10-bit 4K final from a 480p draft, and calls it “the most capable model in the Seedance 2.5 generation” (BytePlus blog, Oct 27, 2026). Our test used Topaz.

What you can send with it

Inputs docs don't list limits; treat as a normal shot

See the takes

Topaz, not the built-in upscaler

We sharpened the 1080p final in Topaz instead: same picture and sound, only sharper, slightly waxy. It came out 1924×1080, so crop it before the edit (8/10).

Creator tips (4)

A 480p draft straight to a 4K final. (BytePlus now offers a 10-bit 4K final with Seedance 2.5 4K; check what your platform does.) — Heavy Pulp
Shoot 480p, then Topaz Astra Creative to 1080p. — Christopher Gwinn
Finish with Topaz Astra 2 Creative. — Christopher Fryant
The model strips grain. Add it back in post. — Jordan Chesney

Seedance Field Guide · @JeremiahDaws · jeremiahdaws.com/seedance

Recipe

Make a character talk

One clean picture of the face, mouth closed or neutral, one face per picture.
A 10–15 s voice clip of that character, with no music under it.
Short lines with times, each line tagged with the voice. (A creator rule of thumb we have not tested: about a third of a second a word.)

Copy this prompt

Rule 1: There is exactly one MAYA, as shown in @Image1.
@Image1 is MAYA: her face, hair and green work jacket.
@Audio1 is MAYA's voice.
Medium close-up, 50mm lens, eye level. MAYA sits in a diner booth, both hands around a coffee cup.
0–3 s: she looks up from the cup. MAYA (@Audio1): "You're late."
3–6 s: she sets the cup down without looking away. MAYA (@Audio1): "Again."
6–10 s: she holds the look; her jaw tightens.
Only these sounds: MAYA's voice, quiet diner chatter. No music. No subtitles.

Options marked Advanced take a little technique, or go beyond the docs. Our tests, not documented means what we show comes from our own tests, not the docs. Everything else is in the official API docs.

Prompt anatomy

A prompt that holds up

Go long: detail helps. A shape that works across all the options:

Rules firstNumbered rules at the very top, before anything else, for what must hold in every shot. Rule 1: There is never a duplicate human face nor a duplicate wardrobe. Rule 2: The environments are always realistic, as shown in image 1. Rule 3: The characters are always felt-fabric puppets, as shown in image 2. Rules like 2 and 3 hold a mixed-media look, such as felt characters in a real-looking world. —

Jordan Chesney

Scene contextLook, period, place, and one line on how it's cut. 1980s diner at closing time, 35mm film look, warm practical lights.

ReferencesName each one once and say what it fixes. @Image1 is MAYA's face and green jacket.

The shot(s)Size, angle, lens, camera move, then the action in order. Give every cut a reason.

PerformanceSmall physical actions beat named emotions. He tries to laugh along. The smile breaks. beats he's sad.

DialogueShort lines, written the way people talk. Listeners keep their mouths shut.

SoundList only the sounds you want. Say "No music" and "No subtitles."

Reference limits

How much you can feed it

Images

Video clips

Audio clips

All files

Two separate time caps: all clips combined up to 30 s, and all audio combined up to 30 s (fal says 30.2 s) (fal, BytePlus docs). You hit those before you hit 10 files. Each input clip: 2–30 s on BytePlus (4–30 s for an edit), 300–6000 px a side, 24–60 fps (BytePlus docs). fal allows 1.8–30.2 s. Images up to 30 MB, resized to 300–6000 px a side. Clips mp4 or mov up to 200 MB. Audio files 2–30 s and up to 15 MB each on BytePlus and Segmind (fal allows 1.8–30.2 s). MP3 on Segmind; MP3 or WAV on BytePlus and fal. Audio-only references are supported (BytePlus, Segmind).

Pro tips

Get the most out of every shot

Keep the face

The same character in every shot.

How: a character sheet as a reference, the same NAME in caps every time, and a clear waist-up picture alongside the sheet.

Just the sounds you want

A clean track, with music only when you ask for it.

How: list "only these sounds" and say "No music" twice. Use music-free voice references. Or ask Seedance to take it out of the take: “Remove the background music from @Video1. Keep the dialogue, the picture and the timing unchanged.” A stem split in post is there too.

Real faces

BytePlus says 2.5 refuses reference images or videos that contain real human faces, not only first frames and source clips (BytePlus docs). Segmind's docs mention first frames.

How: BytePlus offers real-person verification with consent, lets you reuse trusted model outputs (faces its models made under your account), and has a library of preset digital characters. Never use someone's likeness or voice without their permission.

Clear the filter

Keep a harmless shot flowing past the content filter.

How: describe what we see, not graphic outcomes. Change one word at a time to find the trigger. On Segmind, skip_moderation is on by default.

Every run is fresh

Re-running a prompt gives you a new take each time.

How: judge a change over 2–3 takes. To keep a take, use draft-to-final instead of re-rolling. Segmind's seed defaults to a fixed 42; fal notes seeds give fresh takes.

Lines in order

Every line, said by the right speaker, in the right order.

How: short lines, one speaker per beat, and a body action between lines.

Crisp text

From creators, untested here: logos, signs and t-shirt prints can respell themselves shot to shot. In our edit test, the mug kept its logo.

How: plain wardrobe and blank signs. Add text in post.

Tight edits

Change a jacket and only the jacket.

How: name and count everyone, close the scope ("only this changes"), and say what keeps the original timing.

Keep the head in frame

From creators, untested here: mentioning knees or feet can push the frame below the head.

How: "His face is the subject of every shot" works better than a rule against cropping.

Labels and watermarks

Some platforms add a visible watermark and content credentials (C2PA: hidden data in the file that says it was made with AI).

Helps: know your platform's rules before you deliver to a client or a festival.

Good to know

What the docs actually say

You'll hearDrafts expire after 7 days.

The docs sayIt depends where. 14 days on Segmind. On fal, US-region drafts last 7 days. On BytePlus, 14 days for draft IDs made from Oct 28, 2026, 7 for older ones. Finish early. (Segmind, fal, BytePlus docs)

You'll hearStoryboard grids take up to 15 panels.

The docs sayUse as many as you like. 15 is a rule of thumb; fewer, clearer panels read best.

You'll hearClips have roles: structure, depth, camera, performance.

The docs sayThose are tool-specific labels. Officially you name @Video1 and say its job in the prompt, for depth, pose, camera or lip-sync.

You'll hearAudio edit swaps a voice and keeps lip-sync.

The docs sayThat one comes from our own experiments: see The sound card and have fun with it.

You'll hearSeedance can only follow one camera, so split every shot.

The docs sayIt cuts between several shots inside one 30 s generation.

From the creators

The creators, and their tips

Tips for each option are inside its card, under "Creator tips." These apply everywhere. Each one links to the post where they said it, most with the work to watch.

Across every shot

Budget about 5 re-rolls per shot. Some take 10–15; know when to stop. — Christopher Gwinn
Long prompts work: some run past 12,000 characters. Expect trial and error. — Heavy Pulp
2.5 has its own prompt markup. Read ByteDance's official prompt guide. — Christopher Fryant
Stuck on a prompt? Have Claude or Grok rephrase it. — Dr. Dreams
Plan a backup model for a refused shot, and match your grade across both. — Christopher Gwinn
Keep faces original: check for celebrity look-alikes. — David LaChance Jr.
Write long, almost like a novel. Change a few words per take. — Jordan Chesney

Where they split

Who writes the prompt: some write every word and warn against AI-written prompts; others draft with ChatGPT or Claude and edit.
Timecodes: some leave them out; others time every beat to the second.
Rules: some open every prompt with numbered rules; others put the must-haves inside the shot and the sound line.
Start frames: some build every shot from one; others say references look better.
Where they run it: invideo and CapCut, Dreamina, Venice, Luma, Runway. Features and filters differ by platform.

Meet the creators

The people behind the tips in this guide, and a piece of their work. Tap a still to watch it; tap a name to follow them on X.

▶ Watch

A short in 7 prompts

Director. Hybrid live action and AI; posts his full Seedance method.

Kyle Salazar@eattheethos

▶ Watch

The Internet Is Gonna SEND Us

His film “Sweet Land” passed 10M views. Draft-mode evangelist.

Heavy Pulp@heavypulp

▶ Watch

BRO

Makes the ’80s-camp series Bambi the Destroyer. Hosts the Slop & Glory podcast.

Dr. Dreams@DrDreams

▶ Watch

Salty

Photoreal AI films and brand spots. Shares full prompts and how-to threads.

Jordan Chesney@jordandchesney

▶ Watch

Treasure of the Himalayas

Grindhouse Glitch: retro genre shorts. His giallo Amalia hit 165K views in days.

Christopher Gwinn@ChrisGwinnLA

▶ Watch

Pneumatic

27 years in 2D and 3D before AI. Finishes in Topaz.

Christopher Fryant@cfryant

▶ Watch

BLUEBIRD, Ep 4

Red Rainbow, a retro sci-fi horror series with recurring characters.

Aze Alter@AzeAlter

▶ Watch

The Moon Calls

A werewolf tribute. Pushes up to 15 shots into one 30 s clip.

James D Phillips@jdp2oo

▶ Watch

Depth-map choreography

Emmy-winning director. Builds free, open-source film tools.

Sam Wasserman@SamJWasserman

▶ Watch

Directing in the description

Creative director making the series All One.

gen.ericai@gen_ericai

▶ Watch

Clay-render blocking

Builds sets in Blender and directs AI shots from 3D blocking.

alexgnewmedia@alexgnewmedia

▶ Watch

Office test

Composer and sound designer. Scores his own AI shorts.

David LaChance Jr.@DavidLaChanceJr

▶ Watch

Film shots to previz

Turns real film shots into previz to show the model, like a DP.

Josh Daws@JoshDaws

▶ Watch

Can I Make Cone Man More Cinematic?

YouTube breakdowns of Seedance with image models.

David IsraelYouTube

▶ Watch

Wonder Sessions talk

From Google and L’Oréal spots to her own original films.

Olga BaranovaYouTube

Stills and work belong to each creator, shown with credit and links back. Want yours changed or removed? Reach out to @JeremiahDaws.

Test it yourself

One test per option

We ran every test on Mr. Snuggles; what we made is under each one, grouped by the six modes. Each test answers one question. Use your own character sheet, set picture and one strong finished take as the source. Change one thing at a time, and roll 2–3 takes to see the range.

Run T1 and T3 first, since together they show how much references matter. Then T5, then the change tests. Run T11 and T12 last, once you have a winner.

1 · Text to video

Text only

Your character in a simple setting, 8 s, words only.

Q: How close does words-only get? This is the baseline for T3.

What we made: The blocking and every line, from words alone. Add pictures to keep the exact character. Clips →

Ran on: Segmind (#1000195).

2 · First + last frame

First + last frame

Two stills of the same place, a 6 s push between them.

Q: Does it land on the last frame exactly, with no morph?

What we made: A clean push-in, with the last frame a crop of the first. Frames that really differ land too, and it imagines the path between them. For a talking shot, a silent pass then a voice pass gave the best take. Clips →

Ran on: Segmind for the two-stills take (#1000209); Dreamina for the reruns.

3 · References

References

Same prompt as T1, plus the character sheet, a costume picture and the set.

Q: How much better is identity than T1? Do small details hold?

What we made: Host, set and voice all held: the biggest single jump in the tests. To reuse a take, one still pulled from it gave the liveliest result. Clips →

Ran on: Segmind (#1000194); the still-from-a-take tests on Dreamina.

Storyboard

One image of 6 panels, 10 s.

Q: Does it follow panel order and framing? Do the borders stay out?

What we made: Every panel in order, at its framing. A clean-line board renders him as a real felt puppet. Time each shot in the words too. Clips →

Ran on: Segmind (#1000189), then a pencil board and a clean-line board on Dreamina.

Multi-shot

"Shot 1: wide / Shot 2: medium / Shot 3: close", 12 s.

Q: Do the cuts land where written, at the sizes written?

What we made: Words only (#1000196): cuts within ¼ s of where written. A still per cut: every shot matched its still. Clips →

Ran on: Segmind for words only; Dreamina for the still-per-cut rerun.

T13

Clay (Blender)

Build a quick grey blockout of the set in Blender, render the angle you want, and send it as a reference for a fresh shot.

Q: Does it frame the shot like the clay, without turning it grey?

What we made: Yes. It held a brand-new side angle in full colour. Add a set picture to keep your props. Clips →

Ran on: Dreamina.

T15

Depth maps

A depth map per cut (white near, black far) for a fresh 10 s multi-shot. Then the same angles as clay renders.

Q: Does each cut take its map's angle? Does depth beat clay?

What we made: Every angle matched its map. On wild angles, depth maps got all three directions right, the best of the two. Use them on fresh shots. Clips →

Ran on: Dreamina for the fresh shots; Segmind for the re-shoot (#1000212).

T16

Movement

A free stock dance clip, sent three ways: the clip itself, a full-body depth video, and a depth video cropped to the waist.

Q: Does your character copy the moves and stay where he belongs?

What we made: The depth video copied the moves in order and on time. A full-body clip brings a standing dance; a clip cropped to the waist gave us a seated take. Clips →

Ran on: Dreamina.

T17

A look

Black and white noir with blind light, set in words, with a picture of another room, and with a graded picture of his own set. Then a worn 80s VHS look, from a picture used for lighting and filters only, plus words.

Q: Does the look change, with nothing else from the picture appearing?

What we made: A graded picture of his own set brought hard blind light. Words made it black and white. Put them together in the grade for the full noir look. A picture “for lighting and filters only”, plus the look in words, brought the whole 80s VHS look onto his own set. Clips →

Ran on: Dreamina.

4 · Edit

Edit a shot

Your source take: change only one piece of wardrobe. Then swap a prop from a picture.

Q: Does only that change? Watch the other faces.

What we made: Only the tie changed, even the sound stayed. A picture swapped in the exact mug. Clips →

Ran on: Segmind (#1000197, #1000206).

T14

A new place

Your source take, moved to a new location. Once with a picture of the place, once with words only.

Q: Does the performance hold? Is the old set fully gone?

What we made: With a picture, a whole new set and the performance near frame for frame. Words work too; a picture pins the details. Clips →

Ran on: Dreamina.

A new angle

10 s of the source, re-shot from a new camera position.

Q: Do the face, acting and timing match the original?

What we made: A picture of the angle moved the camera furthest, low to a three-quarter view. Asking for four setups in words cut one take into four shot sizes. For a big move, make a fresh shot from a depth map or clay. Clips →

Ran on: Segmind (#1000200, #1000211, #1000212), then Dreamina.

T10

The sound

(a) Remove a music bed. (b) The same lines in another language.

Q: Is the picture untouched? Do the lips match in (b)?

What we made: (a) With the docs' wording, the music came out clean, with the voice and timing intact. (b) Seedance dubbed the lines into Spanish on its own. For his exact voice, ElevenLabs made a Spanish read in my brother's cat voice and Seedance lip-synced him to it. Lip-sync to my own voice file and to a song both worked. Clips →

Ran on: (a) Dreamina. (b) Segmind (#1000204), then Dreamina for the ElevenLabs read and the 3-take retest.

5 · Extend

Extend forward

Last 5 s of the source, plus 5 s more. Try it with Extend, and again with "keep the clip, add seconds."

Q: Which one hides the seam?

What we made: The new part starts almost on the old last frame. A placement picture added a new character exactly where I wanted him, in frame the whole shot. Clips →

Ran on: Segmind (#1000198, #1000208); the new-character reruns on Dreamina.

Extend backward

First 5 s of the source, with a 5 s lead-in.

Q: Does the lead-in end on the source's first frame?

What we made: A lead-in that ends very close to the first frame. Clips →

Ran on: Segmind (#1000199).

6 · Finish

T11

Draft to final

Finalize your best draft at 1080p.

Q: Same take, frame for frame, only sharper?

What we made: The same performance, frame for frame, at 1080p. Clips →

Ran on: Segmind (#1000203).

T12

Upscale

Same take as T11 through an upscaler, side by side.

Q: Can you see the difference on a phone? On a TV?

What we made: We used Topaz: the same take, only sharper. Clips →

Ran on: Topaz, on the Segmind final.

Test log

What worked, and how

Every test on this site was run on one character, Mr. Snuggles, in one set. Mr. Snuggles' voice: my younger brother, Jon Daws, who's been doing this cat for 30 years. I sampled it into ElevenLabs. Here is each thing we set out to do, what we fed it first, and what made it work. The first tests ran on Segmind and the reruns on Dreamina; where a pair crosses platforms, the row says so. The clips, with every take, live on each option's card.

Mode 1

Keep the character

How: Add a picture of him and the set

First try: none, words only
What worked: host + set pictures, Tree voice read

Words alone gave the same blocking and lines. One picture each brought in the exact cat.

Watch the before and after on the Text to video card →

Mode 2

A smooth push-in

How: Make the last frame a crop of the first

First try: medium still → close still
What worked: medium still → the same still cropped tighter

A crop of the same picture gives a clean push-in. (Before on Segmind, after on Dreamina.)

Watch the before and after on the First + last frame card →

Mode 3

Every shot just like the picture

How: Give it a still for each shot

First try: host + set pictures, Dog voice read
What worked: the same, plus a wide, medium and close-up still

The cuts land where written, and with one still per shot, every shot matched. (Before on Segmind, after on Dreamina.)

Watch the before and after on the Multi-shot card →

Mode 3

Borrow only the camera move

How: Strip the sound from the clip

First try: host + set, Dog voice, camera clip with its sound
What worked: the same camera clip, sound stripped

Silent, it copied only the camera move. (Before on Segmind, after on Dreamina.)

Watch the before and after on the Camera move card →

Mode 4

Move the camera on a re-shoot

How: Show it a picture of the angle

First try: source take + host picture, angle in words
What worked: the same, plus a picture of the angle

Asked for a side angle, a picture of the angle moved the camera furthest.

Watch the before and after on the A new angle card →

Mode 3

A brand-new angle

How: Make a fresh shot framed from a clay blockout built in Blender

First try: source take + host + picture of the angle
What worked: a fresh shot: grey clay layout + host + set + Dog voice

A clay still on a fresh shot held the side angle all the way. (Re-shoot on Segmind, clay shot on Dreamina.)

Watch the before and after on the Clay (Blender) card →

Mode 4

A dub in his own voice

How: Send a Spanish read made in ElevenLabs, in my brother's cat voice

First try: source take + host picture
What worked: the same, plus that Spanish read

Seedance spoke correct Spanish on its own. With the ElevenLabs read as the reference, it lip-synced him to his own voice. (Before on Segmind, after on Dreamina.)

Watch the before and after on The sound card →

Mode 5

Add a character

How: A placement picture: a frame of the take with him painted in

First try: source take + host + band picture
What worked: the take + the placement picture

A frame with him already in place put him exactly there, in frame the whole shot. Prompt tweaks shape his actions. (Before on Segmind, after on Dreamina.)

Watch the before and after on the Extend card →

Mode 4

Move the show to a new place

How: Show it a picture of the new place

First try: source take, the new place in words
What worked: the same take + a picture of the diner

With the picture, the whole set became the diner and he played it near frame for frame.

Watch the before and after on the A new place card →

Mode 2

A talking first + last shot

How: Run it silent, then add the voice in a second pass

First try: first + last frames, lines in the prompt (then: both frames as references + his voice)
What worked: first + last frames, silent; then that take + his voice read

Two passes: he leaps in and talks in his own voice.

Watch the before and after on the First + last frame card →

Mode 3

Reuse an earlier take

How: Pull one still from the take

First try: an earlier take as the only reference, with its sound, then sound stripped
What worked: one still from that take + the set + his voice

One still: right words, a clean set, livelier acting.

Watch the before and after on the A still from a take card →

Mode 4

Remove a music bed

How: Use the docs' wording, and render 2–3

First try: a take with an instrumental bed mixed under it, other wording
What worked: “Remove the background music from @Video1. Keep the dialogue, the picture and the timing unchanged.”

With the docs' wording, the bed came out clean, with the voice and timing intact.

Watch the before and after on The sound card →

Mode 3

Dance in his chair

How: Crop it to the waist, roll a few takes for a seated one

First try: a full-body depth video of the dance
What worked: the same clip cropped to the waist

Moves in order and on time. Cropped to the waist, it gave us a seated take. (All on Dreamina.)

Watch the takes on the Movement card →

Mode 3

A noir look

How: A graded picture for the light, words for black and white, then the grade

For the light: a graded picture of his own set
For black and white: the look in words

The picture brought the blind bars. Words made it black and white. Finish the look in the grade. (All on Dreamina.)

Watch the takes on the A look card →

Mode 3

Wild angles

How: Use depth maps for wild angles

First try: three clay renders, one per cut (top, worm's-eye, behind)
What worked: three depth maps of the same angles

Depth maps got all three directions right: top-down, worm's-eye and from behind.

Watch the before and after on the Depth maps card →

What we fed it: every reference, as sent

Same host, same set, same voice reads through every test. Tap a picture for full size.

Pictures

**Host picture**Every test but words-only

**Set picture**Refs, multi-shot, camera clip, clay

**Close still**First + last (two stills)

**Medium, cropped tighter**First + last rerun (true zoom)

**Picture of the new angle**New angle from a picture

**Depth map of that angle**New angle from a depth map

Grey clay layout (Blender) — **Clay blockout (Blender)**Fresh shot framed from clay

**Clay painted into a frame**Clay as a first frame

**Empty desk (first frame)**First + last, frames that differ

**Him at the desk (last frame)**First + last, frames that differ

**Head sheet**References: head + body sheets

**Headless body sheet**References: head + body sheets

**Depth map, side (Blender)**Depth maps: one map, fresh shot

**Depth map, front**Depth maps: three per cut

**Depth map, side**Depth maps: three per cut

**Depth map, close**Depth maps: three per cut

**Clay render, top-down**Clay: three angles

**Clay render, worm's-eye**Clay: three angles

**Clay render, from behind**Clay: three angles

Clips

**Source take**Every edit, extend, re-shoot and dub

**Moving clay clip**Clay as a camera move

Sound

Voice read, Tree bit (15 s)References, edits, extends, re-shoots

Laugh track (15 s)The source take

Voice read, Dog bit (10 s)Multi-shot, camera clip, clay

Spanish read (ElevenLabs, my brother's cat voice)The sound: dub rerun

Sources

Checked against

Last checked against the docs: Nov 13, 2026.