Assembly or process video from a 3D render with Seedance 2.5

To make an assembly or process video with Seedance 2.5, split the process into steps of 4 to 30 seconds, give each step a reference for the structure and another for the look, and join the clips in Timeline in order. ByteDance's launch post shows the idea: a prompt that takes camera work, composition, part positions, model structure, assembly order and motion paths from a clay render, and materials, lighting and colour from a photo, and returns a photoreal car assembly sequence. On Sume you can send those references to seedance-2.5 as image and video references.

What does the vendor example do?

In the launch post (read 2026-10-03) the car-assembly prompt reads the structure from a clay render reference and the look from an image reference. The post lists the clay render among strengthened references, alongside motion and creative references, and describes industrial use such as process training and equipment demonstrations.

It does not promise the assembly order will be correct for your product. The same post says physical plausibility of complex motion is still improving. Treat the output as a visual, and have someone who knows the real procedure review it.

How do you plan the steps?

Write the procedure as numbered steps and give each its own request. Short steps are easier to review: you can reject step 4 without redoing steps 1 to 3. Keep each step to one action, such as "the bracket slides into the frame", and a duration of 4 to 10 seconds. Reserve 30 seconds for a single continuous overview shot at the end.

The Video Router doc says references are accepted by seedance-2.5 and that limits are per model. Keep the count of references modest, name what each one is for in the prompt, and do not combine first/last-frame fields with reference fields in the same request.

How do you join and tighten the clips?

Timeline 1.0 places up to 200 slots in order against an audio spine; use audio.mode: "silence" for a silent process video or a narration file as the spine. Each slot has source_in and duration, so you can use only the best part of a clip. Video trim makes a new MP4 for a [start, end) range if you want a reusable cut.

Run the unbilled POST /v1/timeline-1.0/plan before the render: it returns the duration, segment count and an estimate, and it catches problems in your slot list before they cost anything.

What goes in the prompt for one step?

State the references' jobs first, then the action, then the camera. For example: "Use the render reference for part positions and the camera path. Use the photo for materials and lighting. Step 3 of 8: the side panel slides onto the frame and locks in place. Slow lateral tracking shot, no cuts." The vendor's own prompts use timestamp ranges such as 0 to 4 seconds and 4 to 7 seconds inside one long clip; that works for a short sequence, while separate requests are easier to review and redo.

Avoid asking for on-screen text unless you plan to check it. Add labels and step numbers afterwards with a caption pass, so the words are yours.

How do you keep steps consistent?

Reuse the same look reference in every step so the material and lighting stay the same. Sume has no seed, so a rerun of one step is a different take; accept that and compare the new step against its neighbours before you place it. If a step is close but the camera is wrong, rerun that step with a clearer camera line instead of changing the references.

For a long overview, a single 30-second request is the longest seedance-2.5 clip; for anything longer, join clips. Start at 480p, then rerun only the approved steps at a higher resolution.