Max duration
30s per clip
4–30s per generation; extend for longer cuts
Seedance 2.5 is ByteDance’s next-generation AI video model, built for up to 30-second video generation, advanced multimodal reference control, precise local editing, and more flexible, production-ready creative workflows.
Seedance 2.5 Draft on APIPod is a two-stage workflow. Submit a normal Seedance 2.5 request to get a low-cost 480p draft; once it completes, submit its task ID as drafttaskid to render the 1080p final video. The final video reuses the draft prompt, media, duration, aspect ratio, seed, and audio setting, and runs on the same generation account as the draft.
Seedance 2.5 Video Edit on APIPod pins the upstream omni-reference task type to video editing. Provide the source video (4-30 seconds) plus optional reference images or audios and describe the edit; the output keeps the source timeline, and the aspect ratio is adaptive.
Seedance 2.5 Video Extension fixes the upstream full‑modal reference subtask to video extension: input the source video and specify the continuation direction (forward or backward); outputs 4‑30 seconds with the frame size fixed as adaptive.
Seedance 2.5 Reference to Video Model, Full‑Power Version, comes with a material library, no facial distortion, 100% real‑human pass rate, long‑term stability. Priced at approximately 90% of the official rate. Material and Scene Limitations Text Prompts: Chinese text is recommended to be no more than 500 characters; English text is recommended to be no more than 1000 words. Image Input: Supports jpeg, png, webp, bmp, tiff, gif, heic, heif; single‑image size below 30 MB, aspect ratio (0.4, 2.5), width and height range (300px, 6000px). Image Count: No more than 30 images. Video Input: Single‑video duration [2, 30] s; up to 10 reference videos allowed; total duration of all videos shall not exceed 30 s. Audio Input: Single‑audio duration [2, 30] s; up to 10 reference audio clips allowed; total duration of all audio clips shall not exceed 30 s. Audio cannot be input independently; at least one reference image or video shall be provided simultaneously. 🔸Note🔸: Unit price for with‑video input is lower due to different calculation formulas: No video = Unit Price × Output Volume; With video = Unit Price × (Input Volume + Output Volume).
Seedance 2.5 image to video model, full‑power version, comes with a material library, no facial distortion, 100% real‑person pass rate, long‑term stability. At least one start‑or‑end‑frame image must be uploaded. The price is approximately 10% off the official price.
Seedance 2.5 Text‑to‑Video version supports generating 4‑30‑second videos. Prompt: "It is recommended that Chinese prompts do not exceed 500 characters, and English prompts do not exceed 1000 words."
Official full-capability at 5% off · Seedance 2.5
ByteDance's next-generation audio-video joint generation model — built for 30-second narratives with precise reference control.
Seedance 2.5 doubles the single-clip ceiling to 4–30 seconds and raises the multimodal reference budget to 30 images, 10 videos and 10 audios per request. On APIPod you call the official full-capability model at 5% off list price through one unified API — discount and special-offer routes share the same endpoint, and plain public URLs are reviewed for you automatically.
Max duration
30s per clip
4–30s per generation; extend for longer cuts
Native audio
Audio + video
Dialogue, ambience and music synced to the frame
Reference budget
30 / 10 / 10
Up to 30 images, 10 videos, 10 audios per request
API routes
6 models
t2v / i2v / r2v × discount / special-offer
Same Seedance DNA; twice the runtime and a much larger reference budget.
Seedance 2.5 extends the 2.0 contract instead of replacing it — the model field is the only change most integrations need.
| Spec | Seedance 2.5 | Seedance 2.0 |
|---|---|---|
| Duration per clip | 4–30s (default 4s) | 4–15s (default 5s) |
| Resolution tiers | Discount: 480P / 720P / 1080P; special-offer: 720P only | Up to 4K |
| Reference images | Up to 30 | Up to 9 |
| Reference videos | Up to 10 | Up to 3 |
| Reference audios | Up to 10 | Up to 3 |
| Native audio | Yes — on by default | Yes |
Both generations ship through APIPod's full-capability pipeline: media URLs are reviewed by the asset service before submission, so plain public URLs work in requests. Billing is token-based on successful tasks only — see the pricing page for live rates.
Official list price · 5% off
Seedance 2.5 bills by tokens on the official side: token consumption per second scales with resolution, and a video reference is billed for the reference duration together with the output duration at a lower unit price.
| Resolution | Tokens per second | Without video reference | With video reference |
|---|---|---|---|
| 480P | 9,607.5 tokens/s | $10.7 /M tokens · ≈ $0.103/s | $6.4 /M tokens · ≈ $0.062/s |
| 720P | 21,600 tokens/s | $10.7 /M tokens · ≈ $0.231/s | $6.4 /M tokens · ≈ $0.138/s |
| 1080P | ≈ 38,500 tokens/s | $10.7 /M tokens · ≈ $0.412/s | $6.4 /M tokens · ≈ $0.246/s |
With-video billing counts reference-video duration plus output duration as tokens. Official example — 10s reference + 10s output: 480P $1.23, 720P $2.76 per clip.
List prices from the official Seedance 2.5 pricing page; APIPod's discount tier bills at 5% off. The special-offer routes (seedance-2.5-lite-*) output 720P only — 1080P is not available there. Your actual rates on APIPod are shown on the pricing page.
Longer narratives, better control — the official positioning, delivered through one API.
Write the clip as timed beats — setup, action, payoff — and Seedance 2.5 holds blocking, wardrobe and scene logic across the full take. Extend twice for even longer cuts.
Up to 30 images, 10 videos and 10 audios per request: character identity, product appearance, motion paths and voice tracks each get an explicit job in the prompt.
Dialogue, ambience and music are generated together with the visuals and land on the frame — including invented-language performances and perfectly timed SFX.
Orbits, whip pans, speed ramps, handheld tracking and invisible-cut staging: name the move and the model executes it with cinematic framing.
Swap backgrounds, restyle scenes or remove elements from an existing clip — the output keeps the source timeline and ratio. Fix the intent up front and constraints fail fast at submission, not minutes later.
View docsProlong an existing clip forward or backward with coherent motion and audio — chain into multi-shot narratives beyond the 30-second single-take limit.
View docsseedance-2.5-draft renders a low-cost 480P preview first. When it completes, send its task_id back as draft_task_id — just the model, draft_task_id and output format — and the 1080P final reuses the exact prompt, media, duration and seed. No re-roll, no surprise; one draft can spawn multiple finals.
View docsCommunity cases from the APIPod prompt library — every card shows the actual video and the prompt skeleton that produced it. Copy them, swap in your subject, and ship.
T2V · 30s action
A dialogue-free 30-second aerial battle written as five timed beats — speed ramps, whip pans and match cuts through volumetric storm clouds.
Prompt excerpt
30-second cinematic aerial action spectacle, high-budget feature film, aggressive speed ramps, rapid match cuts, whip pans, rotating camera moves, volumetric storm clouds, ocean spray… [0–5s] dark storm clouds gather above a vast ocean; a battleship cuts through enormous waves; a fighter emerges from the clouds…
R2V · Identity lock
Four labeled reference sheets hold character, wardrobe and vehicle identity through a full-speed anamorphic chase that ends in a rollover explosion.
Prompt excerpt
Preserve the exact identity and wardrobe of the male target @target_character_sheet and the female assassin @assassin_character_sheet. Use @landcruiser_sheet as the identity reference for the black 1990s Toyota Land Cruiser, @kawasaki_ninja_sheet for the green Ninja motorcycle…
R2V · Native audio
Two reference portraits become a blockbuster battlefield exchange performed entirely in an invented alien language — generated audio, generated picture.
Prompt excerpt
Image1 is the alien commander and leader; Image2 is the soldier alien. The commander gives orders to the soldier on a battlefield. A dialogue scene between the two — the speech must be entirely in an invented alien language. A battle scene from a blockbuster film, with cinematic camera moves and angles.
T2V · Multi-scene consistency
One traveler held across eight distinct scenes — city, beach, ocean, market, sunset, hotel — with handheld travel-diary movement and golden-hour light.
Prompt excerpt
Create a cinematic 30-second tropical travel vlog featuring the same 20-year-old East Asian woman with dark hair throughout every scene. Keep her facial identity, hairstyle, natural makeup and body proportions consistent. Authentic handheld travel-diary movement, candid performance, soft golden-hour light…
Run seedance-2.5-draft to validate composition, timing and reference roles cheaply — and test a 5–10s opening before committing to the full 30s.
Start from the community templates in the prompt library; each detail page explains the beat structure so you can rebuild the rhythm for your subject.
Pass the winning draft's task ID as draft_task_id: the 1080P final reuses the same prompt, references, duration and seed, so what you approved is what you get. Need 720P? Rerun the prompt on the t2v / i2v / r2v routes; fix a single failed beat by editing only that time range.
Switch between 2.5 and 2.0 tiers by changing the model field — auth, billing and observability carry over.
Example request
Public model IDs on one endpoint: discount routes seedance-2.5-t2v for text-to-video, seedance-2.5-i2v for first/last-frame animation and seedance-2.5-r2v for multimodal references; special-offer seedance-2.5-lite-* routes bill at special-offer rates (720P only). Two-stage seedance-2.5-draft renders a 480P draft first, then a 1080P final from draft_task_id.
Media URLs are reviewed by the asset service automatically — pass plain public URLs and the pipeline exchanges them for approved asset references. Tasks are asynchronous; poll the task ID until completed.
curl -X POST https://api.apipod.ai/v1/videos/generations \
-H "Authorization: Bearer $APIPOD_API_KEY" \
-H "Content-Type: application/json" \
-d {
"model": "seedance-2.5-t2v",
"prompt": "A cinematic tracking shot through a rain-soaked neon street, realistic motion, synchronized ambient sound.",
"duration": 8,
"resolution": "720p",
"aspect_ratio": "16:9",
"generate_audio": true
}Seedance 2.5 is ByteDance's next-generation audio-video joint generation model, built for 30-second storytelling. It generates video with synchronized native audio, follows timed-beat prompts, and accepts large multimodal reference packs for identity, motion and voice control.
A single generation runs 4–30 seconds (default 4s). For longer narratives, write the clip as timed beats and extend the result — the official guidance allows extending twice for longer cuts.
2.5 doubles the per-clip ceiling from 15s to 30s and raises reference limits from 9 images / 3 videos / 3 audios to 30 / 10 / 10. On resolution, the discount routes output 480P / 720P / 1080P and the special-offer routes (seedance-2.5-lite-*) are locked to 720P, while Seedance 2.0 supports up to 4K. Prompt structure, native audio and the asset-review pipeline are shared.
Yes for 1080P. The discount routes (seedance-2.5-t2v / i2v / r2v) output 480P / 720P / 1080P; the special-offer routes (seedance-2.5-lite-*) are locked to 720P. For 4K masters, use the Seedance 2.0 standard tier (4K now live on APIPod).
Yes. Audio-video joint generation is on by default (generate_audio=true): dialogue, ambience and music are produced together with the visuals. You can disable it per request, and steer voices and sound with reference audio clips on the r2v route.
POST to /v1/videos/generations with model set to seedance-2.5-t2v, seedance-2.5-i2v or seedance-2.5-r2v. The task is asynchronous — poll the returned task ID until status is completed. Media URLs you pass are reviewed by the asset service automatically, so plain public URLs just work.
Yes — via the dedicated seedance-2.5-edit and seedance-2.5-extend models. Both pin the omni-reference sub-task type so parameter constraints are validated at submission time instead of failing asynchronously. Editing output follows the source video (duration and ratio not settable); extension output is 4-30 seconds. Both accept @视频N / @图N tokens in the prompt to address materials explicitly.
Four steps with the same model, seedance-2.5-draft. 1) Submit a normal request at 480p to get a draft, billed at the 480p rate. 2) Poll GET /v1/videos/status/{task_id} until status is completed — that task's data.task_id is your draft_task_id. 3) Submit only model and draft_task_id, plus optional resolution (1080p) and output_format; the prompt, media, duration, aspect ratio, seed and audio setting are reused, and resending them returns an error. 4) Poll the new task for the 1080p final, billed at the 1080p rate for the draft duration. Drafts stay valid for 7 days and both stages run on the same generation account.
Official full-capability model at 5% off, deeper special-offer pricing, one unified API — your first clip is one request away.