Every exhibition trailer begins with assets that already carry a history: installation photographs, artist portraits, catalogue scans, music, and footage supplied under a specific agreement. Interest in
Seedance 3.0 is understandable, but a curator cannot treat a new model name as permission to remix that history. The useful question is whether the production desk can prove what entered the generator, what the model was asked to preserve, and why the final cut was cleared.
SeedVideo currently marks the new version as coming soon and routes actual creation through Seedance 2.0. That live workspace accepts text, images, videos, and audio references in one form. It offers more control than a prompt-only tool, but it also creates a larger rights surface. A beautiful result can still be unusable if one reference was licensed only for print, an artists face changes between shots, or a soundtrack was supplied for internal review rather than publication.
Start With Rights Before Choosing Any Reference
A reference is not merely visual inspiration once it is uploaded to a generation system. It becomes production input. The desk should therefore attach a simple use status to every asset before anyone opens the model: cleared for generation and publication, cleared for internal exploration only, or not cleared. This small distinction prevents an exploratory render from quietly becoming the public trailer.
The platform terms say users retain ownership of uploaded material and, subject to compliance, own generated output. They also make use subject to applicable rights and the policies of underlying model providers. Ownership language does not repair a missing photographer release or expand a museum loan agreement. The institution still needs its own chain of permission.
Separate Artist Approval From Technical File Access
A shared drive link proves that a producer can open a file; it does not prove that the producer may use it as an AI reference. Put approval beside the asset, not in a distant email thread. If a portrait can guide identity but cannot appear as a recognisable frame, say so. If a performance recording may inform movement but not sound, split those permissions as well.
Build One Reference Pack Around One Visual Claim
The current form allows up to nine reference images, three reference videos, and three reference audio files. Video references can total no more than 15 seconds, and audio references share the same 15-second total. Those limits are generous enough to describe a complex idea, yet they are not an invitation to upload every item in the press folder.
A coherent pack should answer one claim: this is the artists appearance, this is the installations spatial rhythm, or this is the camera move the trailer needs. When several files contradict one another, the model has to infer which detail matters. A different hairstyle across portraits, two incompatible lighting schemes, or music with a tempo that fights the reference movement can produce a polished clip that nobody can defend.
Give Every Image a Single Production Job
Label each image by function: identity, costume, object, gallery space, palette, or composition. Remove an image if its job duplicates another weaker source. Nine clean references are not automatically better than three precise ones. The first render test should reveal whether the minimum pack holds the intended identity before more material is added.
Trim Motion References to the Exact Gesture
A 15-second allowance disappears quickly when editors upload long handles around the useful action. Cut a camera reference to the move itself and retain only enough lead-in to understand direction. For performance, preserve the starting pose, transition, and landing. That makes a later failure observable: the model missed the turn, rather than vaguely not matching the video.
Prepare the Prompt as a Shot Contract
The prompt field supports a long instruction, but length should serve control rather than decoration. A museum trailer brief works better when it states subject, action, camera, setting, light, sound, and prohibited changes in that order. It should also name which reference supplies each element. The goal is not literary flourish. It is to create language that an editor, curator, and rights manager can all inspect.
1. Write one sentence describing the clips editorial purpose and intended audience.
Assign each approved reference to identity, movement, space, or sound.
Specify the required action and camera move without stacking optional scenes.
Add rejection conditions for changed identity, invented artwork, unreadable labels, or mismatched audio.
That last line is essential. A model may create plausible material that was never in the archive. If a generated wall label invents a date, or a sculpture grows an extra element between shots, visual plausibility becomes an editorial risk. The cut should be rejected even if the motion feels cinematic.
Match Aspect Ratio to the Exhibition Channel
SeedVideo exposes landscape 16:9, portrait 9:16, square 1:1, standard 4:3, mobile 3:4, ultrawide 21:9, and Auto. It also shows 480p, 720p, and 1080p output choices and durations from four to 15 seconds. These settings belong in the brief before generation, because they change the composition rather than simply resizing it later.
A wide establishing view may hold an installation and visitor movement, while a vertical social cut needs a much tighter subject path. Cropping one into the other after generation can remove the artwork, clip captions, or make the camera move feel accidental. Editors usually save more time by generating toward the delivery frame than by rescuing a mismatched composition in post.
Review Three Frames at Final Display Size
Check the opening frame, the most complex interaction, and the ending frame inside the actual web card or social layout. Compare faces, hands, artwork edges, and any visible text beside the approved source. A clip that looks fine in a large preview may become unreadable in a small announcement card. This check catches a practical failure before the file reaches the exhibition team.
Keep Seedance 3 Claims Out of Approvals
The live page labels Seedance 3 as upcoming, so no curator should write its unannounced resolution, duration, price, or release date into an approval note. The documented model available now is Seedance 2.0. ByteDance describes that version as a unified multimodal audio-video system supporting text, image, audio, and video input, with up to nine images, three videos, and three audio files.
The official release also identifies useful boundaries. Multi-subject consistency, text reproduction, complex edits, multi-person lip matching, and occasional audio distortion still need improvement. Those are not footnotes for the end of production. They are reasons to keep artwork captions out of generated pixels, review every recognisable person, and retain a conventional edit path for critical sound.
Approve the Clip With an Evidence Bundle
The final handoff should include the cleared reference list, prompt version, chosen ratio and duration, generated file, and a short rejection log. The log does not need to archive every unattractive attempt. It should explain why a result never published: identity drift, an invented object, an altered artwork, a rights conflict, or sound that failed the intended cue. That record turns rework into institutional memory rather than another private argument.
SeedVideo fits art teams that want multimodal references in one creation surface and are willing to keep human approval around it. It is less suitable as an unattended publishing route, because aesthetic fluency cannot certify provenance. The strongest use is modest and specific: let the model help stage motion, while the institution remains responsible for rights, identity, factual detail, and the final public meaning.
