Before an AIGC brand film enters production, teams often receive a very “complete” reference pack: a film still for atmosphere, a product photograph for the subject, an architecture image for the world, and several clips for camera rhythm. Everything is compressed into one moodboard. The direction appears aligned—until the first keyframes reveal that everyone meant something different by “match the reference.”
The model may not be the main problem, and a longer prompt may not fix it. The reference pack is being asked to answer four different questions at once: what must remain the same, how the space works, what the world feels like, and how time moves through it. If those jobs are not separated, beautiful frames may still fail to form a consistent film or survive different formats and campaign versions.
This article proposes a four-lane reference system for a brand world: identity anchors, spatial skeletons, material and light, and motion grammar. It is ONCE’s bounded production framework, informed by current public tool documentation and commercial-film practice. It is not a universal validated model, nor a promise that any generation tool will solve consistency automatically.
A style reference is not a brand reference
Official product documentation reviewed on 6 October 2026 already separates several reference functions. Tencent Cloud’s CreateAigcVideoTask documentation accepts multiple reference images for supported models, while its related data structure can distinguish asset and style reference types in applicable workflows. Runway documents reusable references and Brand Kits organised around products, locations, colours and other identity assets. Adobe Firefly separately documents style references and video composition references. Features will change across models and versions, but the production lesson is stable: more references are not automatically better; every reference needs a defined control job.
A red futuristic city image may clearly describe colour, haze and scale, but it cannot verify the ports on the three-quarter rear view of a product. A clean product photograph can lock appearance, but it does not tell a model where a performer enters or how the camera passes a foreground object. If both are summarised as “keep it consistent,” review notes have no precise language.
Public peer portfolios reveal another limitation. They can show finished AIGC brand films, product commercials and concept pieces, but rarely expose how reference assets were assigned, which attempts were rejected, or how assets were reused between versions. A portfolio can establish that a format exists. It cannot replace a project-specific control plan or prove that one workflow is always faster or more effective.
Lane one: identity anchors define what cannot drift
Identity anchors protect the factual subjects in a brand world: products, people, props, packaging, marks and any structure viewers may interpret as real information. They are not selected because they have the strongest mood. They are selected because reviewers can compare against them.
A useful identity set may include approved product photography, final character designs, packaging artwork, 3D files, locked wardrobe and props, and the necessary views of each subject. Divide those assets into two groups:
- Invariants: product proportions, component count, port location, packaging text, identifiable facial features, statutory marks and approved brand marks.
- Controlled variables: pose, shot size, environmental reflection, condensation, partial occlusion and wardrobe combinations that remain within the approved design.
The decisive question is not how many images were supplied. It is whether a reviewer can name the error. “The product feels wrong” is still a taste reaction. “At 2.8 seconds the right-side port has doubled” or “the hairline changes across the reverse angle” gives production a repairable observation.
An identity anchor must also survive the edit. The same person has to remain the same person across shots; props have to continue; and a product entering a fantastic world still needs credible scale. Correctness in one frame is not enough if identity dissolves in the cut.
Lane two: a spatial skeleton defines relationships
A spatial skeleton does not decide how dreamlike the image should feel. It defines the relationship between subject, foreground, ground plane, architecture, light source and camera. It can come from a location photograph, blockout, simple 3D scene, sketch, storyboard or an approved live-action plate.
Many images read as “AI” because relationships fail rather than because textures look synthetic: a person has no weight on the step; a product touches a table without a contact shadow; scale changes incorrectly during a move; or a vertical crop removes both the subject and the spatial cue. Additional style adjectives will not fix an error in spatial structure.
A spatial reference should define at least three things:
- Subject position and scale, including safe areas for horizontal, vertical and subtitled versions.
- Contact, occlusion and shadow relationships between the person or product and the set.
- Permitted camera movement and the relationships that must remain readable after the move.
This lane also clarifies the value of live action. A brand may not need to film the entire environment. It may need a small, credible physical island: the real product, a hand, the tabletop and the key light. Generative imagery can then extend the city beyond a window or enlarge the distant world. When contact, reflection and occlusion come from a controlled plate, the imagined world has somewhere reliable to land.
Lane three: material and light define how the world feels
Material and lighting references are closest to a traditional moodboard, but they still need to move beyond words such as “premium,” “cinematic” or “futuristic.” Specify roughness, reflection spread, black level, key-to-fill relationship, colour temperature, atmosphere and colour proportion.
The same identity anchor can enter very different visual worlds. Soft daylight can make metal feel domestic and approachable. A narrow hard source can emphasise edges and mechanical precision. Low-contrast haze can add depth while making packaging text harder to read. No style is universally right; it must serve the brand proposition without obscuring information that the shot is responsible for carrying.
This is where style most often conflicts with product truth. A colourful mirrored reference may be visually compelling, but if it makes a matte product look chrome, the team cannot defend it simply because the moodboard is attractive. In a product-decision shot, truthful material should lead. In a transitional world-building shot, atmosphere may have greater freedom.
Lane four: motion grammar defines how the film moves through time
Static references can protect subjects and visual tone, but they do not automatically define speed, pause, gaze direction, camera inertia or the causal logic of a transition. Motion grammar needs its own short references: the rhythm of a push or orbit, the start and end of a performance, particle direction, the order of a transformation, and the sync point between sound and visual change.
A motion reference should not be described only as “move like this clip.” Break it into reviewable actions:
- Does the shot begin at rest, or is it already moving?
- Does the subject move first, or does the camera lead?
- Does speed obey physical inertia or a deliberately graphic beat?
- Is the transition triggered by shape, occlusion, light or continuous action?
- Does the final frame need to hold for product reading, or continue into an open ending?
Motion grammar does not mean every shot uses the same move. It means different shots share a temporal logic. A film that travels from reality into imagination may begin with weight and physical inertia, then gradually release movement as the concept world takes over. Generation itself becomes part of the narrative change instead of a random effect applied to every shot.
When references conflict, define the job of the shot
The four lanes will conflict. A character reference may use natural light, the spatial reference may be a cold architectural scene, and the material board may ask for warm grain. A fast orbit may make product geometry difficult to preserve. The solution is not to keep adding images. It is to state what each shot is responsible for.
A hero shot carries recognition and memory, so identity commonly leads while space and light serve the subject. A transition shot carries a change of world; subject accuracy may loosen, but the motion cause and direction into the next shot must remain clear. A version shot has to survive horizontal, vertical and short edits, so safe area, holding time and information legibility may matter more than an elaborate camera move.
There should not be one permanent priority list for an entire film. Write one sentence for every shot: “If the references conflict, this shot protects…” A product close-up may protect geometry and text before environmental colour. A concept wide may protect scale and motion direction while keeping only the subject silhouette. A performance close-up may protect facial identity and emotional continuity while allowing the background to change.
Stress-test three shots before generating the full film
Production does not need to generate the whole treatment first. Three diagnostic shots can expose the weakest reference lane:
- Hero shot: combine identity, space, material and motion in one demanding shot and observe which lane fails first.
- Transition shot: test whether the move from reality to concept space has a cause, rather than merely producing an attractive random morph.
- Version shot: adapt one piece of information from 16:9 to 9:16 and check whether composition, safe area and subject recognition survive without simply cropping the horizontal frame.
Different failures require different routes. Weak environmental colour or poor motion timing can justify another reference or generation pass. A correct subject with a drifting background can often be repaired through masks, compositing or local reconstruction. Repeated errors in product geometry, packaging text, human identity or physical contact should trigger a move to live action, controlled 3D or an approved plate—not more random sampling disguised as quality control.
The ONCE AIGC commercial and live-action hybrid service similarly separates visual development, keyframes, live-action anchors, generated shots and version delivery. The page describes potential scope; it does not claim that every project uses the same tools or deliverables. The public AIGC case directory can also help teams compare the different jobs of brand advertising, concept worlds and lifestyle imagery. Publicly visible work does not support unreported media or sales claims.
The deliverable is a reusable world guide, not one moodboard
Once the hero film is approved, the four reference lanes should not disappear into chat history. A practical handoff retains source and licence status, approved identity anchors, spatial skeletons, material and lighting samples, motion references, keyframes, rejected attempts and their reasons, model and version notes, manual repair locations, format safe areas and final approvals.
For releases in mainland China, generated-content labelling also belongs in delivery review. The Cyberspace Administration of China and three other authorities published the Measures for Labelling Artificial Intelligence-Generated and Synthetic Content in March 2025, effective 1 September 2025. The rules include visible and embedded labelling requirements for applicable service scenarios. Specific obligations depend on the service role, release route and project; this article is not legal advice.
A reusable brand world does not mean “make every future image look like this one.” It means the team knows which relationships must remain, which variables can be explored, and what wins when references conflict. AIGC can then extend an approved world instead of drawing a new identity from scratch every time.
If you are planning an AIGC brand film or a live-action and generative campaign, share the current product assets, brand guide and a small set of failed tests through ONCE’s commercial production and AIGC entry point. We can use three stress-test shots to identify which reference lanes are ready before deciding what is worth generating and what needs to be filmed precisely.
Sources and scope
Reviewed on 6 October 2026: Tencent Cloud’s AIGC video task documentation and reference-image data structure; Runway Gen-4 Image References, reusable references and Brand Kits; Adobe Firefly style and video-composition reference documentation; the Chinese Measures for Labelling Artificial Intelligence-Generated and Synthetic Content; and ONCE’s Chinese and English AIGC service pages and public case directories. The four-lane system, shot priorities and three-shot stress test are ONCE production analysis, not platform promises or a universally validated model.
