Seedance 1.5 Pro API

Seedance 1.5 Pro is ByteDance's first video model that hears what it shows. Where Seedance 1.0 produced silent clips, the new model writes picture and sound in the same pass: dialogue with matching lip movement, footsteps that land on the right frame, rain, engines, music. ByteDance Seed announced it on 16 December 2025 and Volcano Engine, ByteDance's cloud arm, launched it two days later at its FORCE conference.

The headline is speech. ByteDance built the model to deliver lip-synced dialogue in Mandarin and several Chinese dialects, as well as English, Japanese, Korean, Spanish and Indonesian, and to choreograph the camera like a director would, from slow dolly-ins to the vertigo-inducing dolly zoom. In its own tests it beats Google's Veo 3.1 on Chinese speech and both Veo 3.1 and Kling 2.6 on lip synchronisation.

The other headline is speed. A distilled, quantised inference stack makes Seedance 1.5 Pro more than ten times faster end to end, according to ByteDance, which makes sound practical as a default rather than an occasional extra. It runs here as Seedance 1.5 Pro, for text-to-video and image-to-video.

Who makes Seedance and where 1.5 Pro comes from

Seedance comes from ByteDance Seed, the research organisation behind the Doubao models and the Seedream image generators. The 1.5 Pro technical report lists close to three hundred authors under the name Team Seedance. The first generation, Seedance 1.0, arrived in June 2025 in Pro and Lite versions, built around stable motion and clean multi-shot clips, but it had no sound at all.

ByteDance describes the step to 1.5 in simple terms. Seedance 1.0 was about raising the floor, making motion stable enough to trust. Seedance 1.5 Pro "aims higher", at the ceiling of visual impact, and adds audio as a native part of generation rather than an afterthought.

The model carries the identifier Doubao-Seedance-1.5-pro. Individual users can try it in Dreamina, the international version of ByteDance's Jimeng creative app, and in the ModelArk Experience Center, with Doubao and Jimeng integration due within December. Enterprise API access opens on BytePlus, ByteDance's international cloud, on 24 December, while Volcano Engine serves customers in China.

Seedance 1.5 Pro is a closed model. There are no public weights, no parameter count in the technical report and no licence beyond the commercial terms of the platforms that serve it.

What changed from Seedance 1.0

Seedance 1.5 Pro is a rebuild around sound, but the visual side moved too. The concrete changes:

• Native audio and video in one model. Speech, sound effects, ambience and music are generated together with the frames, so a slammed door and its bang share a moment. The same model can still produce silent video when audio is switched off.
• Lip sync across languages and dialects. ByteDance highlights Sichuanese, Cantonese, Shanghainese and Taiwan Mandarin alongside standard Mandarin, plus English, Japanese, Korean, Spanish and Indonesian. BytePlus adds Portuguese to the list. The model matches mouth movement to the number and identity of speakers, which is where earlier audio-video systems often slipped.
• A camera with its own ideas. The report describes what it calls "autonomous camera scheduling": continuous long takes, Hitchcock-style dolly zooms, orbit, arc and tracking shots, scene transitions and colour grading.
• More vivid motion. ByteDance argues that many rivals buy stability by slowing everything down, and tuned 1.5 Pro for expressive action, faces and camera work instead.
• Better story sense. The model fills in missing visual details and writes dialogue that fits the intended tone, so a loose prompt still returns a coherent little scene.
• Much faster inference. Distillation and serving optimisations cut generation time by more than ten times compared with running the full model.

How it works: two branches, one transformer

The technical report describes a dual-branch Diffusion Transformer built on the MMDiT design that also underpins modern image models. One branch handles video, the other audio, and a cross-modal joint module ties the two streams together during generation. ByteDance points to that deep interaction between sound and picture as the reason it keeps a spoken syllable aligned with a lip shape, and an impact sound with the frame in which it happens.

The model was pre-trained on several tasks at once over large mixed-modality datasets, so one network covers text-to-audio-video, image-to-audio-video and plain silent text-to-video and image-to-video. ByteDance built a dedicated data pipeline for it that filters for audio-visual coherence and motion expressiveness, schedules data in a curriculum, and captions both the picture and the soundtrack in professional detail.

Post-training came in two stages: supervised fine-tuning on curated audio-video clips, then reinforcement learning from human feedback with several reward models scoring motion quality, visual aesthetics and audio fidelity. ByteDance says infrastructure work made that RLHF pipeline nearly three times faster to train. For serving, a multi-stage distillation framework reduces the number of denoising steps, and quantisation and parallelism do the rest, giving the end-to-end speed-up of more than ten times "while preserving model performance".

Seedance 1.5 Pro specifications

ByteDance's tests and early reception

ByteDance measures the model on its own SeedVideoBench 1.5, an expanded version of the benchmark built for Seedance 1.0. It adds advertising and micro-drama scenarios and, for the first time, an audio track: speech, singing and non-verbal sounds, voice attributes such as accent and emotion, and sound effects and music. Film directors helped write the criteria, and experts from film, cinematography and design did the rating, both on a 1-to-5 absolute scale and in head-to-head good-same-bad comparisons.

On video, Seedance 1.5 Pro was compared with Kling 2.5, Kling 2.6, Veo 3.1 and Seedance 1.0 Pro. ByteDance reports a large improvement over its predecessor, the lead in instruction following for text-to-video, and strong competitiveness on aesthetics and motion, both in text-to-video and image-to-video. On audio it was compared with Veo 3.1, Wan 2.5, Kling 2.6 and Sora 2. The report claims a consistent win over Veo 3.1 in Chinese dialogue, dialects and monologues, with few dropped or mispronounced syllables, tight alignment of sound effects with on-screen events, and lip sync that matches the number and identity of speakers better than both Veo 3.1 and Kling 2.6. It also concedes a point: Sora 2 delivers more vivid emotional inflection, while Seedance 1.5 Pro is more restrained, which ByteDance frames as an advantage for controlled professional work.

These are in-house numbers, published as charts rather than an independent leaderboard, and ByteDance is candid about the gaps. Its launch post names motion stability, dialogue between several characters and singing as areas that still need work, and says traditional opera performance is "not yet on par with professional" standards.

Early hands-on testing agrees on the headline. Reviewers single out lip sync as the standout feature, with natural facial motion and convincing timing in both monologues and two-person exchanges, and note that the model follows detailed dialogue prompts closely. The complaints are about action: fast, complex movement such as fighting can warp faces and bodies, and the model tends to brighten a deliberately underexposed start frame.

Where Seedance 1.5 Pro is strong

• Talking characters. A character can deliver a line to camera with mouth shapes, intonation and expression that match, which makes the model a natural fit for short dramas, explainers and ad spots.
• Chinese-language content. Mandarin dialogue and regional dialects such as Sichuanese and Cantonese are a particular focus, and the report shows the model handling comic timing and the stylised speech of traditional opera.
• Sound that matches the action. Footsteps, splashes, engines and room tone land with the visible event instead of floating over the clip.
• Camera movement. Tracking, orbit and push-in shots are executed with confidence, and the model adds motion of its own when the prompt leaves room.
• Close-ups. Subtle micro-expressions hold across a shot, so quiet moments without dialogue still read as a performance.
• Restrained delivery. Voices are more measured than Sora 2's, which ByteDance pitches as a fit for controlled professional work such as ads and explainers.

Weaknesses and limits

• Fast, complex action. Fights and crowded, fast movement are where limbs and faces most often distort.
• Several speakers. A two-person exchange can work well, but ByteDance itself lists multi-character dialogue as an area that still needs work.
• Singing. Sung vocals are weaker than speech, so a music video still needs a separate song.
• Exposure drift. A dark, moody start frame may come back brighter than intended.
• Length and resolution. Clips stop at 12 seconds, and Unicorn API returns 720p, which suits social feeds better than a cinema master.
• Closed and in-house evaluation. No weights, no architecture size, and the published comparisons come from ByteDance.

Prompting Seedance 1.5 Pro

Write the prompt in the order a director would brief a crew: who is in the shot, what they do, where, how the camera moves, then what we hear. Put spoken lines in quotation marks and name the language or accent if it matters. Name sound effects explicitly; the model takes them as instructions.

A street food vendor in a busy night market flips noodles in a smoking wok, looks up and says with a grin, "Two minutes, it's worth the wait." Sizzling oil, clattering pans, chatter in the background. Handheld camera, slow push-in.

For camera work, describe the move rather than hoping for it:

Dolly zoom on a lone hiker at the edge of a canyon: the background stretches away while she stays the same size. Wind gusts, a distant hawk call.

For image-to-video, let the first frame carry the look and use the prompt for motion and sound only:

The cat on the windowsill slowly turns its head toward the camera and blinks. Soft rain on the glass, a clock ticking in the room.

Keep each clip to one or two speakers and one main action. When you need a locked-off tripod shot, say "static camera" in the prompt. Turn audio off for silent B-roll you will score yourself.

Length is the other lever. Clips run from 4 to 12 seconds in one-second steps, so match the duration to the line: a short sentence fits comfortably in 4 or 5 seconds, while an exchange between two characters needs room for both lines and a reaction. Pick the aspect ratio for where the clip will live, 9:16 or 9:21 for vertical feeds, 16:9 or 21:9 for widescreen; in image-to-video the clip takes the shape of the start frame instead.

How it compares with other video models on Unicorn

Against its own predecessor the choice is easy. Seedance 1.0 Pro is silent and stops at 5 or 10 seconds, while Seedance 1.5 Pro adds a native soundtrack and any length from 4 to 12 seconds. Seedance 1.0 Pro still makes sense when a 1080p render on Unicorn matters more than audio.

Kling 2.6 is the closest rival with native sound. ByteDance's own tests favour Seedance 1.5 Pro on lip sync, while Kling 2.6 is the pick when its particular look or motion suits the project better.

Kling 2.5 Turbo Pro and Wan 2.5 are strong choices when motion and style matter more than dialogue. Hailuo 2.3 is a silent image-to-video alternative. For anything that has to talk, Seedance 1.5 Pro is one of the most direct ways on Unicorn to get speech and picture from one render.

Getting started

Open the Seedance 1.5 Pro playground to try prompts, toggle audio and pick a length and aspect ratio. For code, create a key on the API keys page and send a job with the model id bytedance/seedance-1.5-pro; the docs explain how to poll for the finished video. Only prompt is required. Set with_audio to true for sound, choose duration from 4 to 12 and aspect_ratio from the seven options. For image-to-video, set mode to i2v and pass the start frame in image; the clip then takes the shape of your image.

Verdict

Seedance 1.5 Pro turns ByteDance's video line from silent footage into short scenes that speak, and it does so quickly enough to make sound the default rather than an afterthought. Its lip sync, dialect range and camera sense are the reasons to use it. Fast action, crowded dialogue and singing are the reasons to check every take. For talking-head ads, micro-dramas and social clips with real dialogue, it is one of the most capable video models available.