Use the reference type documented for the exact workflow you selected. An original recording and a processed mask can be alternative inputs in one template, but that does not make them interchangeable in every model. A blue-and-white appearance also does not, by itself, identify a depth map or establish how a service reads it.
What the published template actually says
Oimi's Storm II template offers the original video or an extracted blue-and-white mask, using Seedance 2.5 multimodal reference mode. It does not describe that mask as a universal depth-input format.
Treat this as a template-specific instruction. Check the current workflow's connected input and any extraction step before preparing files. A screenshot of a blue figure on a white background does not reveal its frame format, processing method or connection to the next node.
This documentation is not evidence that Imagild accepts the same mask input.
Preserve a source-to-control relationship
Keep the original video untouched. For a processed copy, record where it came from, the exact excerpt and any crop or timing change. You should be able to compare corresponding moments without searching through several unnamed exports.
Review the start, a wide movement and the ending in both versions. Ask whether the representation still distinguishes the relevant hands, body direction and overlapping limbs. Do not claim that invisible detail has been retained simply because the silhouette looks clean.
If the processed clip omits a movement needed by your shot, return to the source or revise the intended scope. Color simplicity is not the same as complete action evidence.
Verify the upload field instead of guessing
Runway's Apps guide says accepted inputs vary by app and are specified at its upload field. Apply that checking habit to the workflow you are actually using, without borrowing another service's accepted formats.
| Item to establish | Evidence to look for |
|---|---|
| Input role | Named performance, control or reference field |
| Representation | Explicit instruction for ordinary video or processed data |
| Pairing | Whether the original and mask are alternatives or both required |
| Alignment | Matching excerpt, crop and timing where required |
| Output scope | What the workflow claims to transfer |
A file opening in your player establishes playback compatibility. It does not establish semantic compatibility with a control-input field.
Compare alternatives without changing the whole project
If the documented workflow permits both branches, keep the character image, selected action and requested output scope constant. Test one source branch at a time.
Review motion separately from identity and scenery. A result that avoids the original costume but loses the main gesture is not automatically better. Conversely, a faithful gesture with unwanted source text needs a different correction from missing movement.
Record observations in neutral language: which event appears, which direction changes and which details remain uncertain. Do not turn one attractive example into a general claim that masks are more accurate.
Questions about processed references
Is a blue-and-white mask always a depth map?
No conclusion follows from color alone. Obtain the workflow's actual description of the representation and its expected input.
Should I generate a mask with an unrelated tool?
Only use a conversion that produces the representation required by the selected workflow. A similar-looking visualization may encode different information; verify compatibility before spending on generation.