Seedream 5.0 Pro Seedance 2.5 Workflow: What Actually Ships

The Seedream 5.0 Pro Seedance 2.5 workflow has a scheduling problem: half of it is not callable. Seedream 5.0 Pro shipped as doubao-seedream-5-0-pro-260628 and sits in ByteDance's public model list today^[3]. Seedance 2.5 does not. As of July 25, 2026, Volcano Engine's documentation portal still answers a search for it with a card saying the Doubao video generation model 2.5 is "releasing in July, stay tuned"^[4], and the model list carries no doubao-seedance-2-5 entry at all^[3].

The architecture is still right, so this is a build guide against the models that answer HTTP today. Two phases, one worked shot with the costs accumulating as we go, and a note at each step about what changes when 2.5 opens.

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 the shipping Seedance generation, not 2.5. Duration 4 to 15 seconds, 480p through 4K, mode decided by which media fields you set^[9].
• 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^[9].
• Image references do not cut the per-second video rate. The cheaper row on reAPI's pricing table applies to uploaded reference video^[9]. The saving is that you iterate on a $0.032 still instead of a $1.025 clip.
• One finished shot costs about $1.28 including eight rejected compositions. Eight text-to-video attempts at the same resolution would be $8.20^[7][8].
• 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

This is the part I did not expect to like. Seedream's finished image URL stays valid for 72 hours, and Seedance accepts public HTTP(S) URLs in image_urls^[9]. If your chain runs inside that window, the entire handoff is copying one string from the image task's output into the video task's input. No download, no bucket, no re-upload. 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-2.0-face",
    "prompt": "Slow push-in. The violinist lifts the bow and begins to play; rain streaks catch the streetlight; steam rises behind her",
    "image_urls": ["https://cdn.reapi.ai/...jpg"],
    "resolution": "720p",
    "duration": 5,
    "size": "adaptive"
  }'

Mode routing, duration, resolution

Mode is implicit. A prompt alone is text-to-video; adding image_urls makes the same call image-to-video; a first and last frame interpolates between them; reference videos or audio drive style and sound^[9]. There is no mode parameter to get wrong.

Four fields earn their place in that body. size: "adaptive" inherits the still's aspect ratio instead of re-cropping the composition you just approved. duration accepts 4 to 15 seconds and defaults to 5. resolution defaults to 720p and ranges to 4K. And the -face variant is the one that accepts identifiable real people, which is a one-word change to model rather than a redesign^[9].

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

Fifteen seconds is the ceiling for one generation, so chain instead of stretching. Set return_last_frame: true, take the returned frame URL, and pass it as image_urls on the next segment^[9]. Each segment starts on the exact pixels the previous one ended on, so there is no drift at the seam.

Chaining also buys something a single long generation cannot: segment three is independently re-rollable. A flaw at second twenty-two costs you one segment, not the twenty-one seconds before it. This is the honest substitute for the 30-second single pass that 2.5 advertises, and I expect plenty of production work will keep choosing it after 2.5 ships.

One shot, start to finish

Here is the whole thing with the meter running. Rates are reAPI's as of July 25, 2026^[7][8].

Drop the optional motion test and the same shot lands at $1.281. Compare that with discovering your framing problem in video: eight text-to-video attempts at 720p is $8.20, 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 $0.205; a 1K still costs $0.032. So one second of finished video is worth six draft stills, and a five-second clip is worth thirty-two of them. Every composition question you answer in video instead of image space is answered at roughly thirty times the necessary price.

One correction to make loudly, because it is the most common false claim about this pipeline. reAPI's Seedance pricing table has two rows per resolution, and the cheaper one reads "720P & uploaded videos" at $0.125 per second against $0.205 for the plain row. That cheaper row triggers on an uploaded reference video. A first-frame image, a first/last-frame pair, and reference images all bill on the plain row^[9]. Going image-first buys control and cheap iteration. It does not buy a lower per-second 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 shipping 2.0 model already outputs 480p through 4K with 4K in 10-bit H.265, and the 2.5 promotion page mentions no resolution at all^[5]. Carrying an existing capability forward as a headline is how spec sheets get invented.

FAQ

Can I use Seedance 2.5 through any API today?

No. It has no model ID in ByteDance's model list and the documentation portal describes it as releasing in July with no further detail^[3][4]. Tutorials showing 2.5 API calls are showing a shape, not a working endpoint.

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

Not the per-second rate. reAPI's cheaper reference row applies to uploaded reference video only, so a first-frame image bills at the same rate as text-to-video^[9]. The saving is upstream: $0.032 per rejected composition instead of $1.025.

How many reference images can each phase take?

Seedream 5.0 Pro accepts up to 10, fused into one output^[1]. Seedance 2.0 on reAPI accepts up to 9 images, plus up to 3 reference videos and 3 audio clips, with videos and audio each capped at 15 seconds in total^[9].

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

Not within 72 hours^[9]. 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^[9].

Moving the workflow onto 2.5 when it opens

Build both phases now and the migration is small, because 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 carry forward untouched. What changes is a model string, probably a wider reference array, and a duration ceiling moving from 15 seconds to 30.

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. Volcano Engine. Model list — Ark model IDs for the Seedance and Seedream families. Retrieved July 2026 from volcengine.com/docs/82379/1330310
4. Volcano Engine. Documentation search — Doubao video generation model 2.5, "releasing in July, stay tuned". Retrieved July 2026 from volcengine.com/docs/search?q=Seedance 2.5
5. ModelArk. Doubao Seedance 2.5 — 30-Second long narrative, full-mode reference expansion. Retrieved July 2026 from ark.volcengine.com/promotion?modelName=seedance-2-5
6. 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
7. reAPI. Seedream 5.0 Pro — live pricing table. Retrieved July 2026 from reapi.ai/models/seedream-5-0-pro
8. reAPI. Seedance 2.0 — live pricing table. Retrieved July 2026 from reapi.ai/models/seedance-2-0
9. reAPI. Seedance 2.0 API reference — mode routing, duration and resolution limits, billing, frame chaining. Retrieved July 2026 from reapi.ai/docs/seedance-2-0

Further reading

• reAPI. Seedance 2.5 Release Status: What's Real, What's Keynote. reapi.ai/blog/seedance-2-5-release-status
• reAPI. Seedance 2.0 character consistency guide. reapi.ai/blog/seedance-2-0-character-consistency-guide
• reAPI. Seedance 2.5 model page. reapi.ai/models/seedance-2-5