AUG 7, 2026
Seedance 2.5 on the Dreamina API
August 7, 2026
seedance-2.5 is live on POST /videos in both the US and CA regions. It is a new generation rather than a variant of the Seedance 2.0 family:
| Seedance 2.0 | Seedance 2.5 | |
|---|---|---|
| Duration | 4-15s | 4-30s |
| Resolutions | 720p / 1080p / 4K | 480p (default) / 720p |
| Omni references | 9 images, 3 videos, 3 audio | 30 / 10 / 10, 50 total |
| Reference video length | 2-15.4s | 1.8-30.2s |
| Prompt length | 4,000 | 15,000 |
The headline is the reference-video discount: give Seedance 2.5 a video to work from and the per-second rate drops by roughly a fifth. It is the only Dreamina model whose price depends on what you feed it.
Real human faces and cinematic violence both work now
Our Seedance 2.0 real-faces tutorial documented a workaround: Dreamina’s Seedance 2.0 input filter rejected recognisable human faces, so the portrait had to be shipped with a sparse grid overlaid on the face to drop the classifier below its threshold.
Seedance 2.5 does not need it. Every clip below uses the unaltered original portrait — the same image, no grid. If you built the grid step into a pipeline, you can skip it for this model.
Stylised violence is generated too. The boxing clip below is a bare-knuckle-style impact with blood spraying through the light, and it renders without complaint — cinematic gore is not what this model refuses. What it does refuse is covered at the end.
Text to video
No references at all, 480p, 4 seconds. Human bodies under impact are where video models usually fall apart, so that is what we asked for: a boxing ring, one swinging work lamp, and a spray of sweat and blood suspended in the light while the camera arcs around the moment of contact.
curl — seedance-2.5, 480p, 4s, text only
curl --location 'https://api.useapi.net/v1/dreamina/videos' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="seedance-2.5"' \
--form 'prompt="Extreme slow motion inside a boxing ring at night, one hard work lamp swinging on its cable above the ropes. A boxer takes a hook to the jaw — head snapping sideways, sweat and blood spraying outward in a suspended arc through the light. His opponent follows through, shoulder and back muscles taut, gloves laced tight. A ring of onlookers watches from the shadows. The camera arcs around the moment of impact. Desaturated, harsh contrast, anamorphic flare, heavy film grain."' \
--form 'ratio="16:9"' \
--form 'duration="4"' \
--form 'account="CA:[email protected]"'
Omni: image, video and audio together
The same request we published for Seedance 2.0 on PixVerse — one portrait, one soccer clip, one line of speech — re-run on Seedance 2.5. The prompt drops the woman from @image1 into the scene from @video1 and drives her speech from @audio1.
The three references go in together:
@image1 — the face and identity to drop in. Unaltered, no face grid.
@video1 — the scene, pose, ball, and motion to keep.
@audio1 — the speech to drive (about 2 seconds).
Because a video reference is attached, this one bills at the discounted rate.
curl — seedance-2.5, 480p, 4s, omni image + video + audio
curl --location 'https://api.useapi.net/v1/dreamina/videos' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="seedance-2.5"' \
--form 'prompt="Replace the woman in @video1 with @image1, keeping the exact same scene, soccer ball and movements, and make her say @audio1"' \
--form 'omni_1_imageRef="CA:[email protected]:w457:h819:s122199-uri:tos-..."' \
--form 'omni_1_videoRef="CA:[email protected]:w720:h1366:d5000-vid:v107..."' \
--form 'omni_1_audioRef="CA:[email protected]:d2038-vid:v107..."' \
--form 'ratio="9:16"' \
--form 'duration="4"' \
--form 'account="CA:[email protected]"'
Upload each reference with POST /assets/account first, then pass the returned assetRef values as omni_N_imageRef, omni_N_videoRef and omni_N_audioRef. A reference the prompt never @-mentions is ignored and not billed.
Fifteen seconds in one shot
The 2.0 family stops at 15 seconds. Seedance 2.5 goes to 30, and holds a character across the whole thing from a single reference image — here a full time-freeze scene at 720p with a beginning, a middle and a punchline.
curl — seedance-2.5, 720p, 15s, one image reference
curl --location 'https://api.useapi.net/v1/dreamina/videos' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="seedance-2.5"' \
--form 'prompt="A cinematic continuous shot on a palm-lined boulevard at golden hour. @image1 walks confidently toward camera ... She stops directly in front of him and looks him slowly up and down. She flicks open a lighter, lights her cigarette, exhales a long slow plume of smoke into his frozen face, smirks, and says '\''Still not returning my calls?'\'' ... Keep @image1 identity consistent throughout."' \
--form 'omni_1_imageRef="CA:[email protected]:w457:h819:s122199-uri:tos-..."' \
--form 'ratio="9:16"' \
--form 'duration="15"' \
--form 'resolution="720p"' \
--form 'account="CA:[email protected]"'
What it costs
Cost is rate × billed seconds, and the rate depends on the resolution and on whether you supplied a reference video. All figures in US dollars, using each region’s Advanced plan (CA converted at CAD/USD 0.7167).
| US $/sec | CA $/sec | 4-second clip (US / CA) | |
|---|---|---|---|
| 480p, with a video reference | $0.113 | $0.105 | $0.45 / $0.42 |
| 480p | $0.138 | $0.128 | $0.55 / $0.51 |
| 720p, with a video reference | $0.235 | $0.216 | $0.94 / $0.86 |
| 720p | $0.300 | $0.280 | $1.20 / $1.12 |
The reference video’s own duration is billed on top of the output, so a 4-second clip built from a 5-second reference bills 9 seconds. The discount pays off when the reference is short relative to the output.
Failed generations are not charged.
See the Dreamina cost calculator for the full per-model breakdown and plan pricing in both regions.
One gotcha worth knowing
Seedance 2.5 can reject a finished generation with fail_code 2053 — the web UI phrases it as “the video may contain third-party content”. It fires two to four minutes into generation and judges the output, not the prompt.
It is not about violence, and it is not about real faces — the boxing clip and the portrait above both render fine. It triggers when the result would depict something recognisable. In our testing, prompts that named no person and no property still failed because they described one: a costumed superhero in a crowd, a trademarked landmark, and a man with a “movie-star jaw” surrounded by paparazzi who was “clearly the most famous person on the street”. Removing that last phrase was what made the 15-second clip render.
The rule for prompt authors: it is not enough to avoid naming a property or a person — avoid asking the model to produce one.
Full parameter reference, per-model limits and a live Try-It console: POST /videos. New to the API? Start with Setup Dreamina.
Start using the API
A single $15/month subscription covers every API on useapi.net — one token, all services. You connect up to 3 of your own provider accounts per service, and useapi.net handles the REST API, load balancing, and polling.
Pricing, crypto payment, and volume options — full refund within the first 14 days if you have fewer than 50 successful generations.