AI video model
Run ByteDance Seedance 2.5 in your browser: outputs from 4 to 30 seconds with native synced audio, driven by text, first/last frames, or up to 30 images plus 10 videos and 10 audio clips referenced together in a single prompt. Credits are per second and shown before you generate.
Seedance 2.5 is ByteDance Seed's flagship video model, announced on July 31, 2026 and reachable through our API partner from August 7. The official framing is "one-take creation": a single pass produces up to 30 seconds of finished audio-video instead of a 5-second fragment, and a single request can carry up to 30 images, 10 video clips, and 10 audio tracks as references at once. ByteDance describes the shift as moving "from merely generating a clip to completing a creative work". On CreateVision AI it runs at 480p or 720p, 4–30 seconds, with native synced audio on by default.
What it is genuinely good at, in our own testing and in independent reviews, is performance and continuity. Dialogue-heavy shots hold up — one reviewer found 30-second generations "particularly suited to 'talky' scenes" and carried an entire short film on the model's dramatic pacing. Our own reference-driven clip picked up a location from one photo and a person from another and kept both coherent for the full take. Because the family is days old it has no Artificial Analysis arena rank yet; we re-check the leaderboards monthly and adjust what Ava recommends.
Now the parts that will actually cost you time. In-frame text is unreliable: our kitchen clip was asked for a chalkboard reading "SOUPE DU JOUR" and returned "SOUPE DU JWIN" — if words matter, add them in post. Reviewers report that morphing and decoherence still appear in fast action, that generated voices pick up accents you did not ask for, and that prompts tuned for Seedance 2.0 "don't reliably transfer to 2.5". Coverage of the launch mentions 4K; the API we use tops out at 720p today, and we would rather say so than imply otherwise. It is also the most expensive video model on this platform — for drafts and volume, Seedance 2.0 Mini costs a fifth as much, and Grok Imagine 1.5 less still.
Positioning, duration and reference limits are from the official ByteDance Seed release: Introducing Seedance 2.5 — One-take Creation, Flexible Referencing
“A single generation can now produce up to 30 seconds of video, and up to 50 reference materials can be passed into one clip.”
“Version 2.5 expands the scope of what must be tested — from whether a generated result is usable, to whether it can be brought closer to a finished deliverable.”
“Morphing and decoherence still show up in fast action sequences, even with well-structured prompts, and no amount of upscaling removes it.”
All 7 videos below were generated on CreateVision AI on August 7, 2026 — every one on the first take, with the exact prompt shown, no editing and no cherry-picking. The six 4-second 480p clips averaged 2m 38s (range 1m 37s–3m 40s); the 8-second 720p clip took 5m 13s. Sound on.
A ceramicist in a sunlit workshop looks up from the wheel and speaks warmly to camera: "This one finally came out right." Clay-dusted hands, natural skin texture, shallow depth of field, soft window light, ambient studio room tone.
Subtle living portrait: she blinks, exhales, and turns her head a few degrees toward the window light. Hair strands drift, golden dust motes float, the landscape outside slides past. Keep her exact face, freckles, and hair unchanged.
Vertical food macro: molten chocolate poured over a warm hazelnut cake, glossy ribbon folding slowly, steam curling upward, crumbs catching the light, cozy cafe ambience.
Place the woman from @image1 walking slowly up the stone steps toward the red torii gate seen in @image2, early morning mist, lanterns glowing, her face and hair exactly as in @image1. Handheld documentary framing, birdsong and distant temple bell.
A seamless loop: the camera drifts a few centimetres closer and then settles back exactly where it started, her expression softening and returning. No cuts, no new elements.
Close-up of a chef's hands rapidly julienning a carrot on a wooden board, knife tapping fast, a handwritten chalkboard menu reading "SOUPE DU JOUR" visible behind, busy kitchen sounds.
One continuous cinematic shot that travels through four worlds without a cut: a neon-wet Tokyo alley at night, dissolving into a sunlit Moroccan spice market, into an underwater kelp forest lit by shafts of light, into a snow-covered pine ridge at sunrise. Fluid camera glide, rich colour grade, transitions carried by light rather than edits.
Compare Seedance 2.5 with other video models — same workspace, one click to switch.
The model is ByteDance's; the reason to use it here is everything wrapped around it.
Getting Seedance 2.5 directly means Jimeng or Doubao Pro accounts, or a Volcano Engine key that needs a mainland +86 phone number and business verification. Here it went live for every subscriber the day the API opened — sign in, pick the model, generate. No key, no SDK, no invoicing setup, and the next flagship lands on this same page.
Seedance 2.5 is the most expensive video model we run, so our Ava agent only recommends it when the brief actually needs it — an output longer than 15 seconds, or image, video and audio references combined in one pass. For drafts, social variants and everyday clips Ava routes you to Seedance 2.0 Mini or Grok Imagine 1.5 instead, and says why.
Pricing is per second and shown before you press generate: 110 credits per second at 480p, 230 at 720p — so a 4-second draft is 440 and an 8-second 720p final is 1,840. Clips driven by a reference video bill at a lower rate on input plus output seconds. The same plan also covers Seedance 2.0, Veo, Kling, HappyHorse, GPT Image 2 and Nano Banana Pro, so you can draft cheap and finish here without five vendor accounts.
In our seven-clip launch-day test every prompt completed on the first take: 4-second 480p clips averaged 2m 38s, the 8-second 720p clip took 5m 13s. We verified the audio with ffmpeg rather than trusting the spec sheet — every clip with audio enabled carries a real AAC stereo track, and switching it off produces a file with no audio stream at all. Every generation keeps its prompt and parameters in history, so a setup that works is reproducible next week.
| Model | seedance-2.5, by ByteDance Seed (announced July 31, 2026) |
|---|---|
| Modes | Text-to-video, image-to-video (first frame, or first and last frame), reference-to-video |
| Duration | 4–30 seconds in a single pass |
| Quality tiers | 480p and 720p |
| Aspect ratios | 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, plus adaptive (image-to-video always adapts to your photo) |
| References | Up to 30 images, 10 videos and 10 audio clips per request, 50 assets in total |
| Audio | Native synced audio, on by default — verified as real AAC 32 kHz stereo in our test set; can be switched off |
| Credits | 110 per second at 480p · 230 per second at 720p · lower per-second rate when a reference video is used, billed on input plus output seconds |
| Measured speed | 4s/480p averaged ~2m 38s, 8s/720p ~5m 13s (our 7-clip test, August 2026) |
Create a free account — no API key, no phone verification, no business paperwork. The generator on this page is already set to Seedance 2.5.
Type a prompt for text-to-video, attach one photo as a first frame (or two for a first-and-last-frame move), or open Reference to Video to combine images, clips and audio in the same request.
Pick anything from 4 to 30 seconds at 480p or 720p. Draft at 480p — it is less than half the credits — then rerun the setup you like at 720p. The exact cost appears before you generate.
Short clips come back in two to four minutes. History keeps every prompt and parameter, so re-running a variation is one click rather than a rewrite.
One account, the full image and video generation workflow.
Where Seedance, Veo, Kling and Hailuo actually differ on motion, audio and price.
The previous generation — still the better value for most 4–15 second clips.
If budget is the constraint, start here instead — Seedance 2.5 is our priciest model.
Google's cinematic alternative, and when its camera language wins.
Start free with daily credits. Upgrade for premium models, faster generation, and 4K output.
Perfect for getting started
Billed annually · $240/year
Billed annually · $300/year
No — it is the flagship tier and the most expensive video model on the platform, at 110 credits per second at 480p and 230 at 720p. Every account gets daily credits, and Z Image Turbo (0 credits, images) plus Seedance 2.0 Mini are the cheap ways to lock an idea down before you spend on a Seedance 2.5 final.
Up to 30 seconds in a single pass, in one-second steps from 4 seconds up — double the ceiling of Seedance 2.0. That is the main reason to pick it: no stitching, no visible seam where two clips join. A 30-second 720p render costs 6,900 credits, so most people draft at 480p first.
Use 2.0 unless you need something 2.5 alone can do. Seedance 2.0 covers 4–15 seconds at less than half the per-second cost and is the better everyday choice. Move to 2.5 when the clip has to run longer than 15 seconds, or when you need image, video and audio references combined in one request.
Yes — that is its other headline capability. One request accepts up to 30 images, 10 video clips and 10 audio tracks, 50 assets in total, and you point at them in the prompt with tags like @image1 or @audio1. Requests that include a reference video bill at a lower per-second rate but count the input video length as well as the output.
Yes, and it is on by default at no extra cost. We checked the files rather than the spec sheet: every clip in our test set with audio enabled carried a real AAC 32 kHz stereo track, and turning audio off produced a file with no audio stream at all. Reviewers do note that generated voices can pick up an accent you did not ask for.
In our August 2026 test set, 4-second 480p clips averaged about 2 minutes 38 seconds (fastest 1m 37s, slowest 3m 40s) and an 8-second 720p clip took 5 minutes 13 seconds. Generation is asynchronous — you can keep working and the result lands in your history.
Text inside the frame, mainly: we asked for a chalkboard reading "SOUPE DU JOUR" and got "SOUPE DU JWIN". Reviewers also report morphing in fast action sequences and unpredictable voice casting. And although launch coverage mentions 4K, the API we use caps at 720p today — for a higher-resolution finish you would need a separate upscale pass.
Yes, videos you generate can be used in personal and commercial projects under our terms of service. You remain responsible for prompt content, trademarks, and likeness rights.
