Seedance 2.5 video generator with reference control

What Seedance 2.5 is

Seedance 2.5 is ByteDance Seed's audio-video joint generation model, published on the official Seed model site with its own page at seed.bytedance.com/en/seedance2_5. The one-sentence description on that page is the whole positioning: a next-generation audio-video joint generation model, built for 30-second storytelling with precise reference control and powerful editing capabilities.

It sits in the official Seed model index between Seed2.1 and Seedream 5.0 Pro, alongside SeedRealtime, Seed Audio 1.0 and Seed GR-RL. Its predecessor has its own page too, and the two official descriptions show what changed in the framing: 2.0 was introduced as a unified multimodal audio-video joint generation architecture supporting text, image, audio and video inputs, while 2.5 is pitched at length, reference precision and editing rather than at input breadth.

This page describes the model as its publisher documents it, then describes what this workspace actually exposes, and keeps the two apart. Where a number comes from a third-party platform rather than from ByteDance, the row says so.

The four things the official model page claims

These are the claims on the official Seedance 2.5 page, in its own words and its own headings. Everything else on this page either quotes another source by name or describes this workspace.

Two of those four rows are unusually concrete for a launch page, and both are demonstrated rather than asserted: the reference-video row and the white-model row are each backed by samples whose prompts the page publishes in full.

What the official product pages add

The model page describes the model. ByteDance's own product pages describe the shipping product, and they publish a set of numbers the model page does not. Both are first-party; neither is a reseller.

30 seconds in one generation, and the two extensions

The headline number is the easiest one to verify on this page, because the official samples are timestamped. The narrative sample published under Longer Narratives runs 28 seconds as one shot with no cut, and the page breaks its prompt into six timestamped beats ending at 0-6, 6-8, 8-14, 14-20, 20-23 and 23-28 seconds.

That sample is also the clearest illustration of what a 30-second prompt looks like. The published opening sets the room, the light and the relationship between the two characters. The published beats then carry the dialogue and the camera separately, so the model is not asked to infer a timeline from prose. The full set is reproduced in the prompt section below.

The extension claim is narrower than it first reads. The model page says videos can be extended twice and does not publish how long an extension is. The product page says a beta long-video mode reaches 180 seconds. Neither statement is a promise about this workspace, which exposes durations from 4 to 30 seconds and no extension control.

Prompts you can copy

Every prompt below is published by the source named beside it and is reproduced word for word. Five come from the official Seedance 2.5 page. Where that page truncates its own text with an ellipsis, the excerpt is reproduced with the truncation kept, because the rest of the string is not public.

Two of the five are published in full. The other three are what the official page publishes, truncation included. For the narrative sample that costs little, because all six beats are complete sentences and they carry the direction: period, camera move, dialogue, and the emotional turn.

An old neighborhood barbershop is about to close. It is already dark outside, the street is slightly blue, and the interior is lit with warm yellow light. There are old photos, old barber certificates, and faded customer group photos on the wall. Character setting: a barber and a customer. Their relationship is natural and familiar, like longtime neighborhood friends. Avoid exaggerated acting...... [Around 1250 characters omitted]

0-6s: A wide interior shot slowly pushes toward the barber chair. The old man sits facing the mirror as the barber stands beside him with a comb and small scissors, ready to trim his sideburns. Reflections match the action. Barber: “Just a little touch around the sideburns, and you’re all set.” Old man: “Take your time. I don’t have any big date tonight.”
6-8s: Cut to a medium shot. The barber gently lifts the white hair near the old man’s ear with a comb and trims the sideburns with scissors. His movements are skilled, slow, and steady, with only soft snips.
8-14s: Front medium shot. The barber uses a soft brush to sweep loose hair from the old man’s neck and shoulders: left shoulder, right shoulder, then the back of the neck. The old man looks pleased in the mirror. Barber: “How does that look?” Old man: “Just right. Same as always.”
14-20s: The barber smoothly turns the chair back, removes the cape, and shakes off the loose hair. The old man rises slowly; they hold a slightly longer handshake. Old man: “Will you be open tomorrow?” Barber: “Not tomorrow. Tonight’s the last night.” Old man: “Then thank you, my friend.” 
20-23s: Barber: “Thank you for always coming in.” A slightly wider interior shot. The old man slowly walks to the glass door as the barber watches from the center of the shop. A small bell rings softly; the old man steps into blue dusk. The barber rests a hand on the chair and lowers his head for a second.
23-28s: The barber turns off the lights. Cut outside: he locks the door and walks away. The final shot holds on the dim shop and the still barber chair, streetlights reflected in the glass, like a quiet curtain call, fading slowly to black.

[Video Edit]Change the background to a vineyard in France, replace the coffee in the two people’s hands with red wine glasses, change the girl’s outfit into a French floral wrap dress, and add a thin-strap small bag on her body. Keep everything else unchanged.

[Sound Edit] Remove the background music from this video and add narration that fits the visuals.

Use the white-model reference video as the sole guide for camera movement, pacing, framing, subject motion, and blocking. Combine it with stage keyframes to create a 30-second cinematic 3D animated short. Style: dreamy, warm, fairy-tale, childlike fantasy. Keep character designs consistent with the keyframes; do not alter appearances. Expressions should change with the scene...... [around 1000 characters omitted]

Keep the girl and the moss-covered stairs from the green-screen video, and replace the green screen with an imaginative fairy-tale fantasy background suitable for mossy stairs. Tall ancient trees surround the area, their canopies interweaving into a natural dome. Thick vines hang from above, and the ground is covered with soft moss, ferns, and tiny wildflowers. On both sides of the stairs are ancient stone pillars half-hidden by plants, broken arches, glowing mushrooms, and flickering tiny light particles, like a magical entrance forgotten by time...... [Around 650 characters omitted]

What this workspace exposes

The picker is the authority on what you can submit here, and it is a narrower surface than the official apps. The rows below are read from this site’s own model configuration rather than from a competitor page.

Two rows deserve a plain statement. The audio row is a limitation of this picker, not a statement about the model: Seedance 2.5 is published as an audio-video joint generation model and the official page ships its samples with voice, while this workspace passes no audio parameter at all, so audio behaviour is whatever the service applies by default.

The reference ceilings are the other honest gap. Official and third-party pages both quote figures far above 2 images and 1 clip, and the difference is this site’s request shape rather than the model’s contract. If your shot needs more anchors than the picker accepts, that is a reason to use a different tool for it.

Seedance 2.5 and Seedance 2.0 side by side

Both generations have their own page on this site and their own page on the official Seed site. The first block of rows below is what each publisher says about its own generation, read from the two official pages; the second block is this workspace’s picker, which is why the two blocks can disagree with each other.

The duration row is the strongest single reason to prefer 2.5 here, because it is the same shape of difference the official pages describe: 30 seconds in one generation against a shorter one. The audio row runs the other way, and that is a real trade rather than a detail.

Jobs this generation is suited to

The official product page names three use cases and the model page publishes samples under all three. The list below stays inside what a published sample demonstrates.

• Cinematic storytelling in a single take. The published narrative sample is a 28-second barbershop scene with six beats and spoken dialogue, and it is the model’s own demonstration of the length claim.
• Product and brand films where the scene must be fixed, not re-rolled. The published edit samples change a background, a drink and a wardrobe in one instruction while keeping the rest of the frame.
• Multi-character scenes with a motion reference. The published white-model samples hand the model a previz clip as the sole guide for camera movement, pacing, framing and blocking.
• Pick-ups and fixes on footage you already have. The published sound-edit sample asks for the music to be removed and narration added, which is an audio-only edit on an existing clip.
• Green-screen shots. The published sample keeps a subject and a staircase and replaces only the backing, which is the workflow the official page calls green-screen editing.
• Multi-language presentation and localisation, given the eleven languages the official product page lists.
• Social-format work in vertical and wide frames, since the picker’s seven ratios include 9:16 and 21:9.
• Storyboard and beat drafting, using the official beat structure rather than one long paragraph.

Why other sites quote different numbers

The pages ranking for this model quote figures that this workspace does not offer. None of them is necessarily wrong, and the ones below are worth separating into first-party and third-party.

First, the figures that are first-party. The ceiling of 50 multimodal inputs, 4K output, region-level editing, the 180-second beta mode and eleven named languages all appear on ByteDance product pages, so they are the publisher’s own numbers. This workspace simply exposes a narrower request: 480p to 1080p, 4 to 30 seconds, and reference counts in single digits.

Second, the figures that are not. A claimed 20% improvement in prompt adherence over Seedance 2.0 appears on two third-party platforms, in near-identical wording, and no official page this page consulted states it. Treat it as those platforms’ own measurement. The same applies to credit prices, which are provider-specific and change: OpenArt, Higgsfield and JXP each price this model differently, and the only number that binds here is the quote the workspace prints before you submit.

Third, on 4K. One competitor advertises native 4K and the official product page does say clean 4K output, while this workspace stops at 1080p. If you need 4K, the honest answer is that this page cannot produce it.

• 50 references: first-party, but not what this picker accepts.
• 4K and 180-second beta: first-party on the official product pages, not available here.
• Region-level editing: first-party capability, not exposed as a control in this workspace.
• 20% better prompt adherence: third-party claim only, stated by two platforms and by no official page consulted here.
• Credit prices: provider-specific; the workspace quote is the figure that applies to your task.

How to write a 30-second prompt

Both the official sample and the two third-party platforms that publish guides converge on the same advice, and the official samples show why: the timeline is written out rather than implied.

1. Set the scene first in one or two sentences: place, time of day, light, and the relationship between the characters. The official narrative sample opens exactly this way before any camera instruction appears.
2. Break the seconds into beats and label each one. The published narrative sample uses six ranges and the Peking opera sample uses three, and each range carries one camera move and one action.
3. Keep one camera instruction per beat. The opera sample moves the camera to Xiang Yu, then to Yu Ji, then pulls back for a triangular full-stage tableau, one move per range.
4. Put spoken lines inside the beat where they happen, with the language named. Two published samples name French and Chinese dialogue inside the beat rather than describing it.
5. Attach a reference clip when camera and blocking matter more than prose. The white-model sample tells the model to treat the reference as the sole guide for camera movement, pacing, framing, subject motion and blocking.
6. Say what must not change. The published edit prompt ends with "Keep everything else unchanged", which is the instruction that protects the parts that already work.

How to generate here

Four steps, in the order the workspace asks for them.

1. Pick the mode: text to video, image to video, or reference to video when a clip should drive the camera.
2. Write the prompt, or open one from the list above and copy it.
3. Set duration between 4 and 30 seconds, resolution and one of the seven ratios, then attach references if the mode takes them.
4. Read the quote, sign in when the workspace asks, then submit. If a task fails, read its status before retrying so you do not submit a duplicate.

What this page does not promise

A model page should say where its own claims stop. These are the boundaries for this one.

• No free allowance is promised. Whatever credits or trial grant exist are shown in your account, not here.
• The 30-second and extension figures are ByteDance’s published claims for its own product. No clip on this page was generated here to test them.
• Audio is not a control in this workspace. Seedance 2.5 is published as an audio-video joint generation model and the official samples ship with voice, but the picker here sends no audio parameter, so do not read this page as offering an audio switch.
• This workspace accepts up to 2 reference images and 1 reference clip. Pages quoting 50 inputs are quoting the model’s own published ceiling, which is not this site’s request shape.
• 4K and the 180-second beta long-video mode are not available here; the picker stops at 1080p and 30 seconds.
• The claimed 20% prompt-adherence improvement over Seedance 2.0 is a third-party figure. This page repeats it only as a claim made by two platforms, and ran no comparison of its own.
• Three of the five official prompts on this page are excerpts. The official page truncates its own text with an ellipsis and the remainder is not public, so the truncation is kept rather than reconstructed.
• Every clip in the gallery is published by ByteDance or by a third-party platform named on its card. None of them was generated on this site, and a published prompt is not a guarantee that the same result will reproduce.
• Latency, price and queue behaviour depend on the service at the moment you submit, not on anything stated here.

Where the samples and numbers come from

Every figure and every clip on this page traces back to one of the sources below. Where a value comes from a third-party platform rather than from ByteDance, the row says so.