AI Video Reference Limits: How Many Images, Clips and Audio Files

Reference-to-video lets you hand a model your product, your character or a motion clip. The catch is that the counts differ sharply, and an over-limit request is rejected, not trimmed. These are the limits as documented on 2026-10-03.

The counts

Sume's limits come from its catalog constraints. Vendor limits come from the vendor's page and may be wider than what Sume exposes.

Reading the table

Sume's catalog marks Seedance rows as supporting image, video and audio references, but its docs do not publish counts for them, so the ByteDance figures are the vendor's and should be treated as an upper bound until you test. Kling 3 on Sume has no reference fields; it accepts a first frame and an end frame only.

Google's Veo page ties reference images to the 8-second duration, which also covers 1080p and 4K. Omni's reference videos are limited to three clips of three seconds each, and audio inside them is ignored.

Using them in a request

On POST /v1/videos, input_references carries reference media and frame_images carries first and last frames. If both are sent, frame_images wins and the request is treated as image-to-video. Omni uses 0-based tokens such as <IMAGE_REF_0> and <VIDEO_REF_0> in the prompt to point at a reference.

Practical rule: send the fewest references that pin the look. Ten product photos from one angle add cost without adding information. See the checklist for pinning a model.