How Should I Choose and Prepare Reference Assets for Multi-Reference Video Generation?
Build a cleaner multi-reference video pack by assigning roles, removing conflicts, standardizing assets, and testing each reference.
Overview
Choose references by function, not by quantity. Every asset should answer one question clearly: who is the character, what is the product, which outfit matters, what should the scene look like, or how should motion and audio behave? Redundant or contradictory references make the model resolve conflicts unpredictably.
Build a reference pack
Use sharp, well-exposed assets with minimal occlusion. Include useful angles without mixing incompatible identities, seasons, logos, or colorways. Crop out unrelated subjects and remove embedded instructions or watermarks you do not want reproduced. Name files by role, such as character_front, product_side, and wardrobe_detail.
Then map each file to the endpoint’s supported input field. WaveSpeedAI model schemas vary: some reference-to-video endpoints accept images, some videos, and some a particular combination. Validate every asset against the current endpoint schema; if it does not state a count, format, size, or duration constraint, do not publish or hard-code a value.
Keep a manifest containing asset ID, role, source rights, checksum, crop, and the model test in which it was used. That makes results reproducible and prevents accidental reference substitution.
Run an ablation test
Generate once with the minimum reference set, then add one asset at a time. If adding an outfit image weakens face identity, the conflict is measurable. Remove low-value references rather than compensating with a longer prompt.
Prepare for production reuse
Version the pack and retain the exact files with the job record. A smaller, clean, rights-cleared reference set usually gives the team more control than a folder full of loosely related images.
