Seedance 2.0 video generation with multimodal references

What Seedance 2.0 is

Seedance 2.0 is ByteDance Seed's video creation model, published on 12 February 2026 with its own page at seed.bytedance.com/en/seedance2_0. The one-sentence description on that page is the positioning: a unified multimodal audio-video joint generation architecture that accepts text, image, audio and video input, built for the most comprehensive set of content reference and editing capabilities in the industry.

The launch post names the model's API identity as well. The model page carries a Get API button that resolves to the Volcengine Ark console entry doubao-seedance-2-0, and at launch the model was reachable from Dreamina, also called Jimeng, and from the Doubao app. That identifier is worth knowing, because every site that resells access to this generation describes it in its own words, and the names rarely match.

This page keeps three things apart that are easy to blur. What the publisher states is quoted and attributed. What competing pages claim is quoted and attributed to them, with a note wherever the official page does not corroborate it. What this workspace exposes is read off this site's own configuration, so it can be checked rather than believed.

One timing fact comes before the rest. Seedance 2.0 is not the newest generation in the family. The publisher's model index lists Seedance 2.5, whose own page describes it as built for 30-second storytelling with precise reference control, and that page is newer than the 2.0 page this article is about. Nothing here claims otherwise.

The four claims on the official launch post

These are the four highlights the publisher lists, under its own headings, on the launch post dated 12 February 2026. They are the only capability claims on this page that come from the publisher rather than from a reseller.

Five capability areas the launch post documents in detail

The launch post then walks through five areas and publishes a prompt with each one. The list below is that structure, not a summary written here.

• Complex motion and interaction that follow physical laws. The published example is a pair figure-skating routine with synchronised takeoffs, mid-air spins and decisive landings, described shot by shot.
• Multimodal all-round reference. Composition, camera language, motion rhythm and sound characteristics can all be taken from the inputs, and the post states that a written shot list or storyboard can be referenced directly.
• Instruction adherence, subject consistency and prompt-driven camera planning. The post says the model plans camera language itself and designs the visual presentation from the instruction.
• Two-channel audio with multi-track parallel output. Background music, ambient effects and character voiceover are generated together and aligned to the picture, and the post singles out small foley details such as frosted glass, plush fabric, acrylic and bubble wrap.
• Broad scenario fit. Advertising, explainer, game and previz work are named, with the argument that generation replaces part of the effects and live-action pipeline.

The published input and output boundaries

These are the boundaries the publisher states, placed next to the numbers third-party pages advertise instead. Where the publisher has not published something, the row says so, because a blank is more useful than a guess.

Prompts you can copy

Every prompt below is the text the publisher prints with that capability, reproduced word for word. Nothing here is a rewrite, and the sample clips further down that carry a prompt use the same source text.

Competitive figure skating live performance. Opening with a low-angle shot following the skate blades gliding on ice, with clear details of ice shavings and light reflections. Entering the spin sequence: the male skater makes a minor error due to slight axis deviation, causing a brief disruption in spin rhythm. The female skater quickly adjusts her center of gravity, keeps a calm gaze, and signals "Stay with me" with eye contact, guiding the male skater to regain alignment. This transitions seamlessly into clean, stable lift movements with sharp, consistent body posture. The climax is a synchronized jump combination: upright aerial posture, decisive ice landings, with perfect audio-visual alignment. The female skater wears a deep-blue figure skating dress; the male skater is in competitive athletic attire. The full sequence delivers a complete narrative from tense mistake to calm, successful execution, highlighting the technical skill and mental strength of elite pair figure skating.

A girl hangs laundry gracefully. After finishing, she takes another piece of clothing from the bucket and shakes it vigorously. 2. Supports Multimodal "All-Round Reference" for Greatly Enhanced Creative Freedom Seedance 2.0 supports multimodal all-round reference , allowing combined input of various texts, images, videos, and audio. The model can accurately understand multimodal input content and generate output as instructed by referencing elements including visual composition, camera language, motion rhythm, and sound characteristics from the inputs. It can even directly reference text-based storyboards, significantly boosting creative freedom.

Refer to the shooting script in @Image 1, and draw on the storyboard, shot scale, camera movement, visuals and copy in @Image 1. The character is from @Image 2, the scene is from @Image 3, and the props are from @Image 4. Create a 15-second healing short film. 3.

Extend the video length. The camera follows a man in orange riding a brown horse. He speeds up and gallops to a large tree with orange flowers ahead, breaks off two blossoms from the branch. Then others ride into the frame one after another. The camera pushes in to shoot the man in orange clothes dismounting from the horse, then the camera quickly circles around him. He turns to walk toward a woman in white riding a white horse and presents the flowers to her. Style: Chinese classical lady painting, 3D, cheerful folk music, shadow puppet style, with black, white, and orange as the main color palette. 4.

Immersive first-person hand ASMR video. Close-up shot, under warm, soft lighting. A pair of slender hands gently triggers different objects in sequence: the light scratching of frosted glass, rubbing of plush fabric, gentle tapping on an acrylic board, light pinching of bubble wrap, and lightly running fingers across a wooden comb's teeth. Finger movements are slow and gentle. Pure, natural trigger sounds with no background music; the visual atmosphere is relaxing and healing. 5.

What this workspace exposes

These are the controls this site's video workspace actually renders for this model, read from the workspace configuration rather than from a marketing page. What you can submit is whatever the selector shows at the moment you submit, and the price is resolved by the server before the task is created.

Two rows are worth reading against the publisher's numbers. The workspace caps length at 15 seconds, the same ceiling the launch post states for a single generation. It exposes 480p, 720p and 1080p rather than any 2K or 4K tier, because those are the tiers the service behind it actually quotes.

Where the eight clips on this page come from

Four of the figures below are clips the publisher prints on the official Seedance 2.0 model page, and each one carries the prompt printed beside it. The other four are the cases three competing pages publish under this model's name, reproduced so the difference in what each page can support is visible rather than described.

Comparing the downloaded files by SHA-256 rather than by eye turned up one result worth stating plainly. Two of the forty-three files collected for this page are republished byte for byte across sources. The clip seedance.tv prints under its immersive audio-visual heading is the same file as the first clip on the official Seedance 2.0 page, which sits under the official immersive audio-visual heading. Seedance.tv's first clip is also the same file as the hero clip on seedance-2.studio.

A page that republishes a file is describing the official sample attached to it, not an independent result of its own, and nothing on this page credits a clip to the site that happens to serve it. Every figure names the page it came from, and the evidence file behind this section records each source URL, the SHA-256 of the downloaded file and the SHA-256 and byte count of the published object.

Seedance 2.0 and Seedance 2.5, read from their own pages

Both generations have their own page on the publisher's site, which makes this the one comparison here where both sides come from the same source. It compares two published positions, not two sets of test results.

The practical reading is that 2.0 is the generation whose published emphasis is reference breadth and physical plausibility, with a stated ceiling of 15 seconds, while 2.5 is the one the publisher pitches on length. Anyone whose shot needs to run longer than 15 seconds in one take is looking at the wrong entry in this family, and this workspace will not offer a longer duration for this model either.

Why other pages quote different numbers

The four pages the site owner asked to be surveyed disagree with the publisher, and with each other, on several numbers. Each row below states what that page claims and what the publisher states, so the gap is on the record rather than in a footnote.

None of this is an accusation of bad faith. A reseller page describes what its own product does with a model, and a reseller is free to run a 2K or 4K tier the launch post never mentions. What the table shows is that a number on such a page is a claim about that page, and cannot be used as a specification of the model.

Jobs this generation is suited to

The list below is built from the capability areas and the published examples, not from generic use-case copy. Each item names what makes it a fit.

• Two-person and multi-subject interaction, where earlier models produced physical errors. The publisher's own sample is a figure-skating pair executing a synchronised jump and holding a landing.
• Shots where the camera language is part of the brief. The post states the model plans camera language from the instruction and designs a visual template for it.
• Sequences assembled from several references. A character, a scene, props and a storyboard can be assigned to separate images inside one instruction, which is how the published 15-second healing-film prompt is written.
• Continuation work. The extension feature generates further continuous shots from a prompt, and the published example continues a horse-riding scene rather than starting a new one.
• Targeted repairs and changes to an existing clip, since the editing feature is described as acting on a specified segment, character, action or storyline.
• Sound-led pieces where the audio is the subject, which is what the published ASMR prompt does with frosted glass, plush fabric, acrylic and bubble wrap.
• Commercial and explainer content, where the published argument is cost and turnaround rather than artistic novelty.

How to generate here

The workspace on this page opens with Seedance 2.0 already selected. Four steps, in order, are all it takes.

1. Choose the mode. Text-to-video needs only the prompt. Image-to-video adds one reference image, and reference-to-video adds one reference clip.
2. Write the instruction in the order the shot happens. The publisher's own prompts name the format first, then the camera, then what changes, then the sound.
3. Set length, resolution, aspect ratio and whether to generate audio, then read the quote the server returns. Nothing is charged until the quote is confirmed.
4. Open the result, compare it against the instruction, and download the take you want. Re-submitting with one variable changed is the cheapest way to learn how this model behaves on your own material.

What the publisher itself says is still weak

The launch post closes with an evaluation section that lists failures as well as strengths. Those admissions are more useful than any benchmark chart, and they are quoted rather than rephrased.

• Detail stability, hyper-realism and dynamic vitality still require ongoing refinement.
• Audio distortion happens occasionally.
• Multi-subject consistency, text rendering accuracy and complex editing effects still have room for optimisation.
• The post states that the model is still far from perfect and that flaws remain in its generation results.
• In the audio dimension the post describes a massive boost in expressiveness, and separately notes that dialect, traditional opera and singing instructions improved markedly, which is a statement about relative improvement rather than about accuracy.

Where the samples and numbers come from

Every figure on this page traces back to one of the rows below. Where a number comes from a competing page rather than from the publisher, the table says so, and every clip is attributed to the page it was taken from.