Seedream to Seedance: The Keyframe Handoff Workflow

Text-to-video has a structural problem that no prompt fixes: changing one detail rewrites the whole scene. Adjust the lighting and the character's face shifts. Fix the face and the background moves. Every iteration is a fresh roll.

The workaround is to stop asking one model to do two jobs. Lock the frame first with an image model, then hand that approved frame to a video model as reference. That is the Seedream Seedance workflow, and its value is not higher quality per generation. It is that approval becomes durable.

TL;DR

• Two phases, two models. Seedream 5.0 Pro settles what the scene looks like. Seedance handles what moves.
• The handoff is a reference image, not a re-description. The approved frame goes in as image_urls.
• Seedream takes up to 10 references with anchored edits and multilingual typography at 2K.
• Iterate cheaply on the still, where a regeneration costs cents and seconds rather than dollars and minutes.
• Approve motion before details. Once the camera move is right, remaining fixes are local.
• Some layered-workflow claims do not survive checking. We went through them separately in what actually ships.

Why decoupling works

A video generation is expensive in both senses: it costs more per call than an image, and it takes long enough that a bad result wastes real time. Pushing every creative decision through that loop is why teams report burning most of their schedule on retakes.

Splitting the pipeline changes what each iteration costs.

Phase one is cheap and fast. Composition, lighting, product geometry, typography, brand color. These are still-frame decisions, and settling them on an image model means each attempt costs a few cents and returns in seconds.

Phase two inherits a settled frame. Motion generation starts from something already approved rather than re-deriving the scene from a paragraph. Fewer variables move, so fewer things break.

The economic argument is straightforward: the expensive call should run once, after the cheap calls have removed the uncertainty.

Phase one: settle the frame

Seedream 5.0 Pro is the asset step. What matters for the handoff:

Up to 10 reference images with anchored edits, so a product or character can be pinned rather than described.

Multilingual typography at 2K, preserving exact color and layout, which matters when the frame carries brand text that must survive into the video.

Pixel-budget billing, split at 2.36M pixels. A 16:9 frame at 2048×1152 lands in the cheaper tier, which is the right size for most keyframes anyway. Details in the pricing breakdown.

The output of this phase is not a folder of options. It is one approved frame per shot, plus the prompt that produced it.

curl https://reapi.ai/api/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedream-5-0-pro",
    "prompt": "product hero on a stone counter, soft window light from camera left, brand wordmark legible on the label",
    "image_urls": ["https://example.com/packshot.jpg"],
    "size": "2048x1152"
  }'

Phase two: hand the frame forward

The handoff is the part people get wrong. The instinct is to describe the approved frame again in the video prompt. That re-introduces exactly the ambiguity phase one removed.

Pass the frame itself:

curl https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2-5",
    "prompt": "0-5s: [locked medium] no camera movement, steam rises from the cup. 5-12s: [slow dolly in] 0.3 m/s, product stays centered. Do not add text or logos.",
    "image_urls": ["https://cdn.example.com/approved-keyframe.png"],
    "resolution": "1080p",
    "duration": 12
  }'

Note what the video prompt does and does not contain. It describes motion and constraints. It does not re-describe the lighting, the composition, or the product, because the reference frame already carries those.

That division is the discipline. Anything visible in the keyframe belongs in the keyframe, not in the motion prompt.

What each phase owns

The rows that generate the most confusion are lighting and framing. Both feel like they belong in the video prompt because that is where the final output comes from. Putting them there means re-deciding them on every generation.

Where the savings actually come from

The 60%-faster-iteration claims floating around this workflow are marketing figures, not measurements. The mechanism underneath them is real though, and it is worth stating plainly rather than as a percentage.

A rejected still costs one image generation. A rejected video costs a video generation plus the wait.

An approved still is reusable. The same keyframe can drive several motion variants, different durations, different camera moves, without regenerating the scene.

Local edits stay local. Fixing a highlight on the keyframe does not disturb a camera move you already approved, because the move has not been generated yet.

That third point is the one teams discover late. Sequencing approval so the cheap decisions settle first is not a productivity hack, it is what makes the expensive call worth making.

Where the pipeline breaks

Re-describing the frame in the video prompt. The most common error. It gives the video model permission to reinterpret what was already settled.

Approving a keyframe at the wrong aspect ratio. Generate the still at the aspect ratio the video will use. Cropping a 1:1 keyframe into a 16:9 shot reintroduces framing decisions.

Skipping the still for "simple" shots. Simple shots are exactly where the still is cheapest, so the saving ratio is highest.

Treating the reference as a suggestion. If the output drifts from the keyframe, add explicit constraints tying the video to it rather than rewriting the description.

FAQ

Why use two models instead of one text-to-video call?

Because a text-to-video call re-derives the entire scene each time, so changing one detail changes everything. Settling the frame first makes approval durable across iterations.

What exactly gets passed between the phases?

The approved keyframe as a reference image, plus a motion prompt. The visual description does not get repeated.

Should the video prompt describe the lighting?

No. Anything visible in the keyframe belongs in the keyframe. The motion prompt covers camera path, timing, and constraints.

What aspect ratio should the keyframe be?

The same one the video will use. Generating a still at a different ratio and cropping reintroduces the framing decision you already made.

Can one keyframe drive several videos?

Yes, and that is much of the value. Different durations, camera moves, and cutdowns can share an approved frame without regenerating the scene.

Does this work for multi-shot sequences?

Yes. Approve one keyframe per shot, then generate each shot's motion from its own reference. Consistency across shots comes from the stills sharing references, not from one long prompt.

Are the layered-output claims about this workflow accurate?

Several widely repeated specs do not hold up against the official capability tables. We checked them in what actually ships.

Sequencing approval, not just generation

The Seedream Seedance workflow gets described as a quality technique, and it is not really. Both models produce what they produce whether or not you chain them.

What chaining changes is the order in which decisions become final. Composition, lighting, and brand detail get settled on a surface where being wrong costs cents. Motion gets generated once, against a frame nobody is still arguing about. The pipeline is worth building for that reason alone, and the teams who report the largest gains are the ones who moved approval earlier, not the ones who found a better prompt.

References

1. Volcano Engine. Seedream and Seedance model families. Retrieved July 2026 from volcengine.com

Further reading

• reAPI. Seedream 5.0 Pro Seedance 2.5 workflow: what actually ships. reapi.ai/blog/seedream-5-0-pro-seedance-2-5-workflow
• reAPI. Seedance 2.5 camera control. reapi.ai/blog/seedance-2-5-camera-control
• reAPI. Seedream 5.0 Pro price. reapi.ai/blog/seedream-5-0-pro-price