Seedance 2.5 is live — 30-second cinematic video with native audio & real-person references
All Posts
Seedream 5.0 Pro Seedance 2.5 Workflow: What Actually Ships
2026/07/25

Seedream 5.0 Pro Seedance 2.5 Workflow: What Actually Ships

Build a live Seedream 5.0 Pro to Seedance 2.5 workflow on reAPI. See current model IDs, pricing, image-first iteration, and working code.

The Seedream 5.0 Pro Seedance 2.5 workflow is now callable end to end. Seedream 5.0 Pro runs as doubao-seedream-5-0-pro, and Seedance 2.5 runs as doubao-seedance-2.5-face on reAPI.[6][10] The earlier July 25 snapshot in Volcano Engine's documentation has been superseded by production access.

The architecture remains the same: approve composition cheaply in image space, then spend video credits on motion. The two phases below now use the live 2.5 route and current production rates.

TL;DR

  • Phase 1 is Seedream 5.0 Pro, live now. 1K and 2K tiers, up to 10 reference images, 14 languages of native in-image text, and exactly one image per call[1].
  • Phase 2 is Seedance 2.5, live now. Use doubao-seedance-2.5-face for 4–30-second output at 480p, 720p or 1080p.[10]
  • The handoff needs no storage of your own if your chain runs inside 72 hours: Seedream's output URL is public and Seedance accepts public URLs.[6][10]
  • Image references do not cut the per-second video rate. The cheaper row applies to source video, not a first-frame still.[10] The saving is that you iterate on a $0.032 image instead of a $1.335 five-second 720p clip.
  • One finished 5-second 720p shot costs about $1.59 including eight rejected compositions. Eight direct video attempts at the same resolution would cost $10.68.[7][11]
  • Three of the specs circulating about this pipeline are wrong, including the resolution tiers and the transparent-PNG layer export.

Phase 1: the keyframe that gets approved

The job of this phase is to produce one still that a human signs off on, as cheaply as possible. Everything else follows from that.

Which tier to draft at

Seedream 5.0 Pro has two resolution tiers and no more: 1K and 2K, selected on reAPI with quality: "basic" or quality: "high"[1][6]. At 16:9 that is 1424×800 and 2816×1584, inside a total-pixel envelope of 921,600 to 4,624,220 and an aspect range from 1:16 to 16:1[1][2].

Draft at 1K. It costs $0.032 against $0.063, and 1424×800 is already above 720p, so it is a perfectly adequate video reference on its own[7]. Re-render the approved frame at 2K only if the still itself ships as a poster or thumbnail.

curl https://reapi.ai/api/v1/images/generations \
  -H "Authorization: Bearer $REAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedream-5-0-pro",
    "prompt": "Wide shot of a violinist on a rain-slicked Berlin street at dusk, sodium streetlights, shallow depth of field, poster headline NACHTMUSIK in the upper left",
    "aspect_ratio": "16:9",
    "quality": "basic"
  }'

Submission is async. Poll GET /api/v1/tasks/{id} until status is completed, then read output.image_urls[0]. Polling is free[6].

Prompt a frame, not a scene

The image prompt describes a static composition: subject, framing, lighting, materials, and any copy you want rendered in the image. Seedream renders in-image text well enough to skip a typesetting pass, and ByteDance's own technique for that is to put the exact words in double quotes[1].

What does not belong here is motion. Save the camera move for phase two, because describing it now produces a still that looks like a frame grab from a pan.

What to lock before you animate

Three things are cheap to fix in image space and expensive in video: framing, wardrobe, and light direction. Get all three signed off at 1K before a single video credit moves. If your workflow has a human reviewer, this is the gate they should be standing at.

One capability to know about and not expect on reAPI: Seedream 5.0 Pro is the only model in its family with interactive editing, where you mark a region on an input image or write <point> / <bbox> tags into the prompt and it edits inside that region only[1]. reAPI's surface accepts prompt, reference images, aspect ratio and quality tier, and nothing else, so region-level repair is not available there[6].

Phase 2: the motion pass

The handoff is one URL

Seedream's finished image URL stays valid for 72 hours, and Seedance accepts public HTTP(S) URLs in image_with_roles for a first-frame handoff.[6][10] If your chain runs inside that window, the handoff is one string from the image task's output into the video task's input. Mirror the asset yourself only if it needs to outlive three days.

curl https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer $REAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2.5-face",
    "prompt": "Slow push-in. The violinist lifts the bow and begins to play; rain streaks catch the streetlight; steam rises behind her",
    "image_with_roles": [
      { "url": "https://cdn.reapi.ai/...jpg", "role": "first_frame" }
    ],
    "resolution": "720p",
    "duration": 5,
    "size": "adaptive"
  }'

Mode routing, duration, resolution

Mode is implicit. A prompt alone is text-to-video; image_urls supplies reference images; image_with_roles defines a first frame, last frame, or both; reference videos or audio drive style and sound.[10] There is no mode parameter to set.

Four fields earn their place in that body. size: "adaptive" inherits the still's aspect ratio instead of re-cropping the approved composition. duration accepts 4 to 30 seconds or -1 for automatic duration. resolution is limited to 480p or 720p. The production model ID includes -face and accepts real-person references subject to automated review.[10]

Write this prompt about change, not about the scene. Camera move, what enters, what leaves, what the light does. Re-describing the composition here fights the reference image you just paid to approve, and the model splits the difference badly.

Getting past fifteen seconds

Thirty seconds is the ceiling for one Seedance 2.5 generation. For longer sequences, set return_last_frame: true, take the returned frame URL, and pass it as image_with_roles with role first_frame on the next segment.[10]

Chaining also buys something a single long generation cannot: each segment is independently rerollable. A flaw after the first 30 seconds costs one segment rather than the complete sequence.

One shot, start to finish

Here is the whole thing with the meter running. Video rates are reAPI's as of August 17, 2026.[7][11]

StepCallCostRunning total
Draft the frame, attempt 1–81K image × 8$0.032 each$0.256
Test the motionSeedance 2.5, 480p × 5s$0.118589/s$0.849
Final renderSeedance 2.5, 720p × 5s$0.266824/s$2.184

Drop the optional motion test and the same shot lands at $1.591. Compare that with discovering the framing problem in video: eight 5-second 720p attempts cost $10.68, and you would still be holding a clip nobody approved as a frame.

The ratio underneath is the thing to remember. A 720p second costs $0.266824; a 1K still costs $0.032. One second of finished video is worth more than eight draft stills, and a five-second clip is worth roughly forty-two of them. Composition is still much cheaper to settle before animation.

One correction matters because it is the most common false claim about this pipeline. The lower Seedance 2.5 rate—$0.160094/s at 720p instead of $0.266824/s—triggers on source video, not a first-frame image. It also bills the greater of output plus rounded-up source-video seconds or ceil(5 / 3 × output seconds).[10][11] Going image-first buys control and cheap iteration, not the source-video rate.

Three specs in the layered-workflow story that don't hold up

"Layer separation returns transparent PNGs for subject, background, props, and text." ByteDance's capability table has exactly one line touching this, in the output-count row for 5.0 Pro: "supports generating a single image or multiple layers"[1]. No documented parameter requests layers, no response field returns them, and no example appears anywhere in the Seedream tutorials. The same table marks group-image output unsupported on this model, and reAPI returns exactly one image per call[6]. Plan for stills.

"Resolution tiers are 1.5K and 2K." They are 1K and 2K[1]. 1.5K appears in neither the model comparison table, the resolution mapping table, nor the pixel-range spec.

"Seedance 2.5 brings native 4K at 10-bit." The live 2.5 route accepts 480p, 720p and (since August 2026) 1080p — still no 4K; Seedance 2.0 remains the family option for 4K.[10]

FAQ

Can I use Seedance 2.5 through an API today?

Yes. On reAPI, submit doubao-seedance-2.5-face to the video-generation endpoint. The model page and API reference publish the current prices, request limits, and billing behavior.[10][11]

Does feeding a Seedream image into Seedance lower the video cost?

Not the per-second rate. reAPI's cheaper row applies to source video only, so a first-frame image bills at the same rate as text-to-video.[10] The saving is upstream: $0.032 per rejected composition instead of $1.335 for a 5-second 720p video attempt.

How many reference images can each phase take?

Seedream 5.0 Pro accepts up to 10, fused into one output.[1] Seedance 2.5 on reAPI accepts up to 30 images, 10 videos, and 10 audio clips, subject to the documented duration and combination constraints.[10]

Do I need to re-host the image between phases?

Not within 72 hours.[6][10] Base64 and data: URIs are rejected, so anything from local disk needs a public URL first.

Which resolution should the keyframe be?

Draft at 1K, deliver at 2K. A 1K 16:9 still is 1424×800, already above 720p and adequate as a video reference; 2K is 2816×1584[2].

What happens if a generation fails?

Charges are placed on submit and refunded automatically when a task ends failed, and polling never costs credits.[10][11]

Running the live workflow on 2.5

The expensive parts of this integration are not model-specific. Async submit and poll, task-lifecycle handling, storage for approved frames, the retry and refund path, and the review gate deciding whether a still deserves to be animated all apply to the live 2.5 workflow.

Ignore the version number and build the split. The reason the Seedream 5.0 Pro Seedance 2.5 workflow is worth adopting is not that either model is new. It is that deciding your composition costs three cents and deciding your motion costs a dollar, and any pipeline that puts those two decisions in the same API call is paying video prices to answer image questions.

References

  1. Volcano Engine. Doubao Seedream 5.0 Pro guide — capability matrix, interactive editing, resolution tiers, reference limits. Retrieved July 2026 from volcengine.com/docs/82379/2582774
  2. Volcano Engine. Image generation guide — Seedream resolution-to-pixel mapping and model comparison. Retrieved July 2026 from volcengine.com/docs/82379/1824121
  3. reAPI. Seedream 5.0 Pro API reference — request body, tiers, one-image-per-call, retention. Retrieved July 2026 from reapi.ai/docs/seedream-5-0-pro
  4. reAPI. Seedream 5.0 Pro — live pricing table. Retrieved July 2026 from reapi.ai/models/seedream-5-0-pro
  5. reAPI. doubao-seedance-2.5-face — model ID, request schema, reference limits, and source-video billing. Retrieved August 17, 2026 from reapi.ai/docs/seedance-2-5
  6. reAPI. Seedance 2.5 — live production pricing. Retrieved August 17, 2026 from reapi.ai/models/seedance-2-5

Further reading

All Posts

More Posts