The generation interface will accept almost any image you feed it. That is not the same as the image being useful.
I have lost count of the times a team dropped a mixed folder of campaign stills, lifestyle frames, and old product shots into a reference slot and then wondered why the model kept drifting on logo placement or surface response. The model was doing what it was asked. The starting point was simply the wrong shape for the job.
Seedance 2.5 can condition on a large set of images, short video clips, and audio. That capacity is powerful when the references are chosen with the downstream handoff in mind. It is expensive when they are not. The difference between a clean white-box turn and a busy lifestyle plate is not aesthetic preference. It is a decision about what kind of control and cleanup debt you are willing to carry into edit, compositing, and client review.
This piece is the practical decision framework I use when a team is choosing the starting point for a commercial test. It is not a ranking of which input type is “best.” It is a map of what each common starting point solves and what it pushes downstream.
Three Common Starting Points and What They Actually Buy
White-Box and Controlled Product References
A clean white-box or neutral-lighting turn of the product remains the most reliable foundation for most packaged-goods and product-hero work I see.
What it solves:
Product geometry and logo placement stay consistent across takes when the reference set is complete
Surface response and material behavior are easier to lock
Later local adjustments (small rotations, logo emphasis) have a clearer baseline
Legal and brand review have an easier time confirming the product itself is correct
What it pushes downstream:
Environment, lighting, and atmosphere still have to be generated or added
The take can feel isolated if the brief requires the product to live inside a specific world
Reflection and contact shadow behavior may need finishing attention once the product is placed in a real plate
I treat a proper white-box package like a camera body: if the product itself is wrong, everything built on top of it inherits the error. For a thirty-second commercial beat that must keep the product recognizable and correctly oriented, I still start here more often than anywhere else.
Green-Screen and Clean-Plate Inputs
Green-screen or clean-plate references are useful when the team already knows the product or talent will be separated and re-composited.
What it solves:
Clear separation intent from the beginning
Easier integration into existing live-action or CGI plates later
More predictable edge behavior when the generation is treated as an element rather than a finished shot
What it pushes downstream:
Spill, edge quality, and motion of the generated element still have to hold up under the final composite
Lighting direction and intensity must be matched carefully or the element will never sit correctly
The generation stage can look deceptively clean while the real work is deferred to the compositor
I have seen teams celebrate a clean green-screen generation only to discover in the composite that the motion or the contact with the ground plane fights every attempt to integrate it. The starting point was correct for separation; the handoff into finishing was under-planned.
Lifestyle, Environment, and Full-Scene References
Lifestyle stills, location plates, and full-scene references are tempting because they look closer to the final commercial image.
What they solve:
Faster exploration of mood, lighting, and environmental interaction
Useful when the brief is more about world and less about precise product geometry
Helpful for early internal discussions about tone
What they push downstream:
Product consistency becomes harder to maintain if the product is not also strongly referenced
Logo and label details can drift or become secondary to the overall scene
Client notes that require precise product changes often force larger regenerations
Cleanup debt rises because the model is solving for many variables at once
I use full-scene references when the primary risk is mood and environmental fit, not product fidelity. When the product itself is the legal and brand critical element, I keep lifestyle material in a secondary role or introduce it only after the product baseline is locked.
The Decision Rule I Actually Use
Before any reference is uploaded I ask one question:
What is the most expensive thing that could go wrong after the first client note?
If the most expensive failure is an incorrect or inconsistent product, I start with white-box and controlled product references.
If the most expensive failure is integration into an existing plate, I start with green-screen or clean-plate intent.
If the most expensive failure is mood or environmental mismatch, I allow stronger lifestyle and scene references—while still protecting the product with a clean secondary set.
The rule is deliberately conservative. It is easier to add atmosphere later than to recover a product that has drifted across an entire variation set.
How I Build the Reference Package in Practice
For a typical product-focused thirty-second test the package looks like this:
Primary product set
Front, three-quarter, and any critical detail angles under controlled lighting
Current approved logo and label version only
Short motion reference if the product has mechanical, liquid, or condensation behavior that must stay consistent
Secondary environment or lighting set (optional at first generation)
One or two stills that define the intended light direction and color temperature
Clean plates only if compositing is already planned
Explicit exclusions
Old logo versions
Lifestyle frames that contain competing products or unapproved talent
Any asset whose rights status is unclear
I label every file. The habit is the same one I use for workshop bins and cable drawers: if I cannot tell at a glance what the file is and whether it is current, it does not belong in the active package.
Seedance 2.5 can accept a large number of references. That does not mean every available image should be included. Extra references that pull in conflicting information simply increase the chance the model averages toward something no one actually wants.
A Short Story From the Desk
Two months ago a team I was advising arrived with a beautiful lifestyle plate of their product sitting on a sunlit table. The plate looked like the final ad. They wanted to use it as the primary reference for a thirty-second generation. I asked them to run a parallel test with a clean white-box set of the same product.
The lifestyle-first takes were more atmospheric. The product logo, however, softened and shifted slightly across the variation set. The white-box-first takes were cleaner on the product and required a second pass to place the product into a matching environment. When the simulated client note arrived—“make sure the logo reads clearly at the hold”—the white-box path absorbed the note with a local adjustment. The lifestyle path required a new generation cycle.
The dog was unimpressed by either result. Morgan looked at the side-by-side and said the same thing she says about project scopes: the version that looks finished earliest is not always the version that survives the first change request. She was right again.
What to Do This Week
Pick one short commercial brief you already understand. Build two small reference packages for it: one white-box / controlled product set, and one lifestyle or full-scene set. Generate a tightly limited number of takes from each under the same prompt and settings. Then apply one realistic revision note to the best take from each path.
Document which path absorbed the note more cleanly and where cleanup or regeneration became necessary. That single comparison will tell you more about the right starting point for your next real job than any general recommendation.
The model will use whatever you give it. The handoff quality is decided by whether you gave it the right starting point for the work that comes after generation.
If it cannot survive the handoff, it is not a workflow yet.
No letters yet — be the first to write.