Seedance — ByteDance's Seed-team video environment — matters because of its direction, not its novelty: text, image, audio, and video references as ingredients in one workflow. This lab tests whether that multimodal control makes your specific job easier. Recheck interface, access, and pricing at lab time; availability differs by region.

Prerequisites

• 59.1 brief (one short job, 8–12 seconds per shot) and 59.4 continuity facts.
• A reference bundle: approved product/character image, one style/location image, optional audio or motion reference where the current surface supports it.
• A 59.A test card: Seedance runs on the shared brief so it is compared on a real job, never treated as exotic or automatically superior.

Workshop path

1. Choose the brief; assemble the bundle. Name every file (refs/product-front.jpg, refs/style-frame.jpg, refs/audio-bed.mp3). Note anything the current interface does not accept. 2. Write one concise shot instruction in motion anatomy form: subject + action + environment + camera + duration/pace + constraints. Attach the bundle, not adjectives. 3. Generate the first short shot. A few seconds only. Save settings, time, and cost. 4. Inspect reference fidelity + motion. Freeze frames: does the product/character match the bundle? Does the camera do the one requested move? Is on-screen text clean or (better) absent? 5. Generate a continuing second shot that deliberately extends shot one — same subject, same light, next beat. Use reference/keyframe carries where supported. 6. Compare. Same brief, one alternate platform (your 59.A fallback). Score both on the 11-dimension card.

The questions that decide

• Did reference control make this project easier — fewer retries, less drift?
• Is the system available where you work, on your account, at a cost this job allows?
• Does it provide the audio/video/editing control this job needs (dialogue, extension, export ratios)?
• Does it outperform the fallback on *this shot's* hardest constraint — and by enough to switch?

If the answer is "better motion, worse product fidelity," say so and assign each tool its job. That split decision is a valid, professional outcome.

Worked mini-example (adapt to your brief)

Suppose the job is the bakery proof shot: a locked close-up of the scored loaf, steam rising, morning window light from the left. The bundle is refs/product-front.jpg (label and crust authoritative) plus refs/style-frame.jpg (navy/green grade). Shot one's direction names the loaf, the steam action, the counter environment, the locked camera, four seconds, and the constraint "keep loaf shape, crust color, and label position exact; no in-frame text." Shot two continues: same loaf, same light, baker's hand entering from frame right with flour — one new element only. If the hand melts or the label slides, the fix is narrower (shorter clip, simpler action, stronger reference weight), never a brand-new concept. Log both attempts; the failed version with its cause teaches the next production what the winner alone cannot.

What to save

Bundle files with names, both shot directions verbatim, settings/versions/cost, winner files, freeze-frame fidelity notes (label, color, shape, hands), the two-shot assembly, and the comparison paragraph against your alternate. Mark anything untested as "not tested" — gaps honestly labeled beat gaps papered over.

Verify fast

Mute test on the two-shot sequence: same world, same subject, readable proof? Then read the bundle filenames against the frames — every must-preserve fact visible? Log scores, screenshots, cost, and the recheck date. No invented results: untested claims stay marked "not tested."

Common failure: changing the concept after a weak first shot instead of narrowing one variable (shorter clip, simpler action, stronger reference weight). Review gate: drafts stay unpublished until facts, permission for every reference face/voice/brand, captions/legibility, and the continuity facts pass a named owner's check.

Face-drift fixes (when the person stops looking like the person)

Faces fail first. Run this ladder before reconcepting:

1. Shorten to 3–4 seconds. Long generations drift identity; short ones hold it. Prove the face holds at 3 seconds before asking for 8. 2. Lock the camera and simplify the action. One head turn *or* one expression change per shot — never both. No walking, no hand-to-face, no profile-to-frontal spins. 3. Strengthen the reference. Frontal, evenly lit portrait on a plain background beats a stylish angled photo. If the tool takes multiple references, add a second angle of the *same* person, same day, same lighting. 4. Freeze-compare. Screenshot frame 1 vs the reference side by side: eyes, nose-to-mouth distance, hairstyle edge, clothing color. Name the drifting feature out loud, then constrain exactly that ("keep dark bob cut exact to shoulders, no longer hair"). 5. Split the shot. Face close-up (identity) and hand/action shot (motion) as two clips cut together — never demand the model hold identity *and* complex motion in one generation.