AI Video Reference Packs: A Time-Coded Procurement Framework

Brands should stop treating a prompt and a folder of images as a production-ready brief. TikTok Symphony already offers Reference to Video using images, video, audio and text. Its Dreamina Seedance 2.5 update raised the stated ceiling to 30 seconds, up to 50 multimodal references and instructions tied to specific timestamps. That does not merely create room for more inputs. It creates a new procurement object: a reference pack that says which asset controls which moment, what must remain invariant and how the shot will be approved.

Fifty references can also create fifty contradictions. If a brand uploads competing pack shots, unlabeled motion examples and rights-unclear talent footage, a more capable model does not resolve the governance problem. It makes the failure harder to diagnose. The prompt is now only one field inside a larger, auditable production specification.

What TikTok has confirmed—and what it has not

TikTok's June 2026 help documentation lists Text to video, Image to video and Reference to video as Symphony Creative Studio modes. Reference to Video can combine multiple images, videos and audio with a text prompt. The same guidance recommends generating complex multi-scene concepts as separate scenes and combining them in an editor. A one-pass finished commercial is not the platform's default advice for every job. (TikTok Help)

On August 3, 2026, TikTok said Dreamina Seedance 2.5 could generate videos up to 30 seconds, accept up to 50 image, video and audio references, and follow timestamp-specific creative direction. On October 7, TikTok World Canada described Reference to Video as letting advertisers call for exact images or products at specific moments. (TikTok for Business) (TikTok Newsroom)

The boundaries matter. TikTok described 2.5 as rolling out to select paid advertisers in select markets, not as universally available. Its consistency claims are vendor statements, not independent quality tests. TikTok also said the initial 2.5 rollout blocks uploaded references containing real faces and applies AI labels, invisible watermarking and C2PA metadata. Before scoping a campaign, verify the account, market, model and input permissions actually available.

More references do not automatically mean more control

A generation model sees conditions, not the brand team's implicit agreement. Four conflicts are especially expensive:

product references show different colors, controls, packaging versions or lens distortion without identifying the master identity asset;
motion references demonstrate an attractive action that the real product cannot perform;
character, environment, lighting and camera references are pooled without assigning each one to a scene or time range;
an asset is owned, but not cleared for paid media, generative alteration, identifiable-person synthesis or every target territory.

Input count measures volume. The useful output is the number of seconds that meet every declared product, brand, evidence and rights constraint. A still image can show appearance; it cannot by itself prove how a latch closes, how a material reacts or how an interface changes.

The ONCE TIMECODE reference pack

ONCE recommends seven fields for every shot unit. Together, they turn reference assets into a purchasable specification.

T — Target

State the placement, aspect ratio, audience, duration and job of the shot. Is it three-second product recognition, usage context or functional proof? Without a target, a visually polished generation may still have no decision value.

I — Identity

Name the exact product model, geometry, materials, colors, packaging version and any character identity. Choose one master reference. Label other angles as supporting evidence, not equally authoritative alternatives.

M — Moment

Assign each reference to a time range. For example: 00:00–00:02 uses the front pack shot and slow push-in; 00:02–00:05 uses the hand-action reference and side geometry. Timestamps prevent every asset from competing to control the full sequence.

E — Evidence

Separate appearance references, motion references and claim evidence. Mechanical behavior, safety use, material response or comparative performance should be supported by verified footage, engineering demonstrations or approved documentation.

C — Constraints

List non-negotiables and forbidden changes: do not add ports, change button count, distort label text or show an unapproved use case. Translate “on brand” into conditions a reviewer can actually test.

O — Ownership

Record source, licensor, territory, term, paid-media rights, generative-modification rights, identifiable-person consent and expiry handling. A platform's generation permission does not complete the brand's own rights chain.

D — Decision

Define the approver, frames to inspect, acceptable tolerance, maximum iteration count and the trigger for compositing or live action. “Make it more premium” is not a traceable stop rule.

This extends the same principle ONCE applies to AI previs: approve what a shot must prove before choosing the technology that produces it. See AI Previsualization for Commercials: What Brands Should Approve Before Production.

Route each shot instead of forcing one workflow

Route 1: direct Reference to Video

Use it for a single scene and action, low identity risk, no critical readable text and no precision product proof. Atmospheric hero moves, background variation and conceptual demonstrations can fit here when reference assets agree and an error would not change a product fact.

Route 2: generate shot by shot, then edit

Use it for multiple beats, multiple scenes or stricter product continuity. Each shot gets its own TIMECODE sub-pack; editing then controls rhythm, audio, captions, color and market versions. This also follows TikTok's own recommendation to generate complex scenes separately and combine them.

Route 3: live action or hybrid production

Use it for regulated claims, readable interfaces and packaging, exact mechanisms, real material behavior, identifiable people or safety demonstrations. AI can handle previs, environment extensions, transitions and variants. The frames carrying proof remain captured or composited under tighter control.

The routing question is simple: if this shot is wrong, is it merely less similar, or does it change what a buyer believes about the product? The second category should not rely on direct generation alone.

Price approved, reference-constrained seconds

To compare suppliers and production routes, ONCE recommends a fully loaded project metric:

Fully loaded cost per approved reference-constrained shot-second = (reference capture and preparation + rights clearance + generation credits and retries + product/identity QA + editing and compositing + localization + provenance and archiving) ÷ seconds that meet every declared constraint and are approved for the target market.

This is not an industry benchmark or a media-performance metric. It prevents two procurement illusions: a cheap generation means a cheap project, and more files mean more usable assets. If twenty outputs contain only four seconds that pass geometry, brand and rights checks, the denominator is four seconds.

Contracts should also separate “generated successfully” from “accepted for delivery.” The first means a file exists. The second means the shot can legally and accurately run in the specified placement, market and term.

Start with one SKU and three diagnostic shots

Do not begin by asking for a 30-second ad in one generation. Choose one SKU and three shots: product recognition, contextual use and functional proof. Build a TIMECODE sub-pack for each, then pre-route it to direct generation, shot-level compositing or hybrid/live production.

Log only what helps diagnose the system: reference-set version, model and settings, retry reason, failed constraint and final approved seconds. The record should reveal whether failure came from the model, contradictory references, missing evidence or an ambiguous acceptance rule.

Teams without 2.5 access do not need to wait. TIMECODE is a production-governance method that works with storyboards, previs, earlier models and manual post. When the model changes, the brand retains structured references and decision history rather than a prompt that quickly becomes obsolete.

The decision for brand leaders

The structural value of Reference to Video is not that brands can throw more files into a model. It is that “which evidence constrains which moment” can become part of the generation task. Teams that package references with timestamps, rights and acceptance criteria can treat the model as a replaceable production layer. Teams that buy prompts and output counts will continue discovering consistency, compliance and rework risk at the end.

The confirmed facts are TikTok's stated inputs, duration, timestamp controls and rollout limits. ONCE's analysis is the TIMECODE procurement framework, the three-route decision and the cost denominator. The next useful step is not to upload fifty references and hope. It is to decide why every reference exists, which second it controls and what result ends the retry loop.