Seedance 2.0 and 2.0 Fast API

Seedance 2.0 is ByteDance's new flagship video model, and the one that dominated talk about AI video in February 2026. It writes picture and sound together in one pass, takes text, images, video clips and audio clips as input in the same request, and produces clips of 4 to 15 seconds that can cut between several shots on their own. Seedance 2.0 Fast is its accelerated sibling, built for teams that care more about turnaround than the last increment of polish.

The headline is not one feature but the combination. Seedance 2.0 can be handed a character sheet, a clip whose camera move it should copy, a voice sample and a written storyboard, and asked to assemble a scene from all of them. On the public Arena video leaderboards in early April it sat at number one for both text-to-video and image-to-video, at 720p, ahead of rivals rendering at 1080p.

It arrives with baggage. The launch set off a public copyright fight with the major Hollywood studios, and ByteDance has since tightened what the model will accept, especially photos of real people. Both tiers run on Unicorn API as Seedance 2.0 and Seedance 2.0 Fast.

Who makes Seedance and how 2.0 arrived

Seedance comes from ByteDance Seed, the research team behind the company's Doubao assistant, the Seedream image models and the Seed-VL vision-language models. The first Seedance shipped in June 2025, and Seedance 1.5 Pro added audio generated jointly with video at the end of that year. The 2.0 build carries the identifier doubao-seedance-2-0-260128, a date code pointing to late January.

Seedance 2.0 surfaced in ByteDance's Jimeng app in China in early February, went viral within days, and was officially launched on 12 February 2026 through Jimeng (called Dreamina outside China), the Doubao app and Volcano Engine's Ark experience centre. A global rollout through CapCut was announced at the same time. International developers had to wait: BytePlus, ByteDance's international cloud, opened a public beta of the API on its ModelArk platform on 14 April, alongside an enterprise launch with partners such as Runway and Higgsfield. BytePlus says the service is not offered in the United States through its own platform.

Seedance 2.0 is a closed model. ByteDance has not released weights, a parameter count or training details, and access runs through its apps and APIs under its commercial terms and acceptable-use policy.

What changed from Seedance 1.5 Pro

ByteDance's technical report, published on 15 April, describes 2.0 as a move from short clips with limited control to "highly controllable video synthesis natively supporting diverse control signals". In practice that means:

• Four input types in one model. Text, images, video and audio all feed a single unified architecture. A request can carry up to 9 reference images, 3 video clips and 3 audio clips. References can define a subject, a motion, a camera move, a visual effect, a style or a voice.
• Editing and extension. The model can change a specific character, action or plot element in an existing clip, and continue footage forwards or backwards on the timeline. Neither was possible in 1.5 Pro.
• Longer clips. Output runs from 4 to 15 seconds at a native 480p or 720p.
• Multi-shot direction. The model plans cuts and camera angles within one generation, and can take a text storyboard as its script.
• Better sound. Audio is now dual-channel, with background music, ambient effects and dialogue laid as separate layers, timed to the action. ByteDance reports large gains in lip sync, English speech, singing and Chinese dialects.
• Much steadier motion. On ByteDance's own SeedVideoBench 2.0, text-to-video motion quality rose from 2.39 for Seedance 1.5 to 3.75 on a 1 to 5 scale, the largest jump of any dimension, with an average gain of 0.86 across all six. Physical feedback, natural phenomena and intense sports motion each improved by more than 1.5 points.

The report also claims much better text in video for 2.0, with short on-screen text rising from 2.00 to 3.57 on prompt following. Independent testers are less impressed on that point, as below.

Seedance 2.0 Fast: same model family, built for turnaround

ByteDance describes Seedance 2.0 Fast as "an accelerated variant of Seedance 2.0 designed to boost generation speed for low-latency scenarios". It keeps the same modes, the same reference limits, the same 4 to 15-second range and native audio, and renders at the same 480p and 720p. ByteDance has not published how much faster it is or how the speed-up was achieved.

Because the two tiers share an interface, a prompt and its references can move from Fast to the standard model without changes. That makes Fast a natural place to settle framing, timing and dialogue before committing a final render to Seedance 2.0, where the extra polish shows most in busy scenes and close-up dialogue.

How it works, as far as ByteDance says

The technical report is light on internals. It calls the architecture "unified, highly efficient, and large-scale" and says audio and video are generated jointly rather than sound being added afterwards, but gives no parameter count, no description of the network and no information about training data.

What is documented is the interface. An adaptive aspect ratio follows the input image, and an optional web search step for text-to-video lets the model look things up before it generates.

Seedance 2.0 specifications

Benchmarks and early reception

The most independent number is the Arena leaderboard, where users vote blind between two models' outputs for the same prompt. As of 8 April, Dreamina Seedance 2.0 at 720p ranked first on text-to-video with an Elo score of 1450, 79 points ahead of Google's Veo 3.1 with audio at 1080p, and first on image-to-video at 1449, 29 points ahead of xAI's Grok Imagine Video at 720p.

ByteDance's own SeedVideoBench 2.0, which mixes automated metrics with blind expert ratings, tells the same story in more detail. On text-to-video, Seedance 2.0 scored 3.75 on motion quality, 3.43 on prompt following, 3.67 on aesthetics and 3.75 on audio-visual sync, ahead of Kling 3.0, Sora 2 Pro, Veo 3.1 and Kling 2.6 on every dimension. The gap is widest on audio: Seedance 2.0 cleared 3.5 on all three audio measures while most rivals stayed below 2.9. On image-to-video it led all six dimensions, though Kling 3.0 came within 0.13 points on preserving the source image.

The public reaction was louder than any table. Clips spread worldwide, Elon Musk commented "It's happening fast", and the screenwriter Rhett Reese wrote "It's likely over for us." Hollywood answered with lawyers: Disney sent a cease-and-desist letter on 13 February, Paramount followed, and the Motion Picture Association accused ByteDance of launching "without meaningful safeguards against infringement". On 16 February ByteDance said it respects intellectual property and would strengthen its safeguards. Separately, after a Chinese tech reviewer found that a photo of his face alone produced a near-identical copy of his voice, ByteDance suspended that behaviour and stopped accepting realistic human photos and videos as reference subjects.

Early hands-on testing is positive with clear caveats. One reviewer who ran 500 generations found single-character social clips and image-to-video shots usable after one or two retries, called the lip sync "tight", and found multi-shot prompts dependable. The same test found fine hand work, exact on-screen wording and complex multi-character shots weak, and motion coherence slipping in clips longer than about 10 seconds. A second tester praised character consistency from reference images but found the model still "took liberties" with timing and framing when copying a reference video.

Where Seedance 2.0 is strong

• Motion that holds together. Fights, sport, dancing and physical interactions show fewer warped limbs and impossible movements than earlier models.
• Sound that belongs to the picture. Dialogue, footsteps, ambience and music arrive in the same render, timed to the action, in stereo.
• Reference-driven scenes. Characters, products, camera moves and voices can be supplied rather than described, so a series of clips looks like one production.
• Editing rhythm. Multi-shot generations cut like edited footage; editing rhythm was one of its highest-scoring categories.
• Image-to-video fidelity. A first frame keeps its look, including unusual art styles, while the subject moves naturally.

Weaknesses and limits

• 720p is the ceiling. Fine for social feeds, short of a broadcast or cinema master.
• Text is still a gamble. Benchmarks improved, but testers found signs and labels approximate. Add lettering in editing.
• Several speakers at once. ByteDance itself lists lip-sync errors in multi-speaker scenes, along with occasional audio distortion and noise.
• Fine detail and hands. Threading, buttoning and handling small objects remain unreliable, and ByteDance notes minor deformations and high-frequency visual noise.
• Long clips drift. The last seconds of a 15-second clip are where coherence and lip sync most often slip.
• Strict content rules. Realistic faces of real people, celebrities and well-known characters are filtered, so some reference images will be refused.
• Closed and lightly documented. No weights, no architecture details, and most published evaluations come from ByteDance.

Prompting Seedance 2.0

Write prompts like a short shot list: subject, action, setting, camera, then sound. Put spoken lines in quotation marks so the model treats them as dialogue, and name the sounds you want, since audio follows the prompt closely. Keep each shot to one main subject when you can.

A baker pulls a tray of bread from a wood-fired oven, sparks rising, then turns to the camera and says, "First batch is always the best." Crackling fire, a tray scraping on brick.

For multi-shot clips, number the shots and give each its own framing:

Shot 1: wide shot of a fishing boat leaving harbour at dawn. Shot 2: close-up of the old captain's hands on the wheel. Shot 3: low angle as gulls lift off the mast. Engine hum, waves, gulls calling.

For reference work, refer to inputs by their order and say what each one controls:

Use image 1 as the character and video 1 for the camera movement. She walks through a quiet greenhouse and stops to smell a white flower. Soft footsteps, rain on the glass roof.

Draft at 480p or on Fast, and re-render only the takes worth keeping at 720p on the standard model. Leave lettering out of the prompt, and if a scene needs three or more people talking, split it into separate shots.

How it compares with other models

Kling 3.0 is the closest rival. It came within reach of Seedance 2.0 on image preservation in ByteDance's tests, offers higher-resolution modes, and is the safer pick when the delivery format needs more than 720p. Seedance 2.0 is ahead on motion, audio and multimodal references.

Seedance 1.5 Pro remains the simpler option in the family: it also generates sound with the picture, but it is limited to 12 seconds, with no video or audio references, no editing and noticeably weaker motion. Wan 2.6 and Grok Imagine Video both offer 1080p, and Grok Imagine runs to 30 seconds, but on Unicorn neither takes image, video and audio references together in one request the way Seedance does.

Between the two tiers: use Seedance 2.0 for final renders, busy scenes and close-up dialogue, and Seedance 2.0 Fast for drafts, volume and anything where speed matters more than the last bit of detail.

Getting started

Open the Seedance 2.0 playground or the Seedance 2.0 Fast playground to set length, resolution and aspect ratio and attach frames or references. For code, create a key on the API keys page and send a job with the model id bytedance/seedance-2.0 or bytedance/seedance-2.0-fast; the docs explain how to collect the finished video. Only prompt is required. Set mode to i2v and pass image for a first frame, with an optional end_image; references go in reference_image_urls, reference_video_urls and reference_audio_urls. with_audio switches the soundtrack on or off, and return_last_frame returns the clip's final frame, which can serve as the first frame of the next clip.

Verdict

Seedance 2.0 is one of the most complete video models on the market: steady motion, sound that belongs to the scene, multi-shot direction and a reference toolkit few rivals match in one request. Its limits are real, from the 720p ceiling to shaky lettering and crowded dialogue, and its launch raised copyright questions that ByteDance is still answering with filters. For short ads, social clips and reference-driven scenes it sets the bar, and the Fast tier makes iterating on that bar quicker.