AI Video Reference Asset Preflight Checklist

Editorial owner: Genflow Editorial · report a factual correction
Research and product review completed: September 8, 2026

Before an AI video run, give every reference asset one primary job, verify that the selected model and mode accept that job, and remove sources that contradict the approved truth. Record file health, duration and count budgets, provenance, likeness consent, and what may or may not change. Start with the smallest pack that can express the brief; add another asset only when it fixes an observable gap.

This is a generation-input checklist. Use the image-to-video prompt workflow after a source image is approved, the prompt troubleshooting log after an output fails, and the shot continuity checklist after adjacent clips exist.

Build the Asset Contract first

Create one row per proposed reference:

Field	What to record
Asset ID	Stable name, version, and checksum
Primary role	Subject truth, product truth, first frame, last frame, style, composition, action, performance, voice, music, or sound effect
Must preserve	The observable details the output cannot change
May change	Framing, environment, wardrobe, pace, or another explicitly flexible element
Model and mode	The exact Genflow model path intended to consume it
Input budget	Count, size, format, per-file duration, and total-duration allowance
Technical pass	Resolution, artifacts, codec, legibility, audio clarity, and trim range
Rights pass	Source, creator, license, territory, term, likeness consent, and approver
Conflict	Any disagreement with another source-of-truth asset
Decision	Include, replace, trim, transform, or exclude

The contract is an editorial record, not a native Genflow feature. Keep it beside the brief and generation receipts so a reviewer can reconstruct what the model was asked to preserve.

Use ROLE-CAP for the preflight

R — Role: give each asset one primary job

“Reference” is not a single control. A portrait may define identity, a product photograph may define package geometry, a first frame may define the opening composition, and a driving video may define motion. A music track and a voice sample are also different jobs.

Write one primary role on each row even when an asset contains more information. MiniMax's H3 examples explicitly assign different inputs to motion, character, and audio roles. Runway's Act-Two similarly separates a performance video from the character source. Those are useful patterns, but they are product-specific capabilities—not proof that every model can interpret the same pack.

If a role cannot be stated as an observable requirement, the asset probably does not belong in the first run.

O — One master: appoint a single source for every fixed fact

Choose one master for each fact that cannot conflict: face, product version, logo geometry, garment, palette, screen direction, spoken language, or action timing. Other references may reinforce that master, but they cannot silently redefine it.

For example, do not submit three product photographs showing different cap designs and ask the model to “keep the product accurate.” Select the approved SKU image as product truth. Mark the others as lighting or angle inspiration only, or exclude them. Do the same when two portraits disagree on hair, age, wardrobe, or accessories.

Conflict review should end with a decision, not a warning. Replace, crop, relabel, or remove the losing asset.

L — Limits: route the pack to the current model and mode

Pick the generation path before approving the pack. A first-frame input, a subject reference, a composition video, a performance driver, and a video-to-video source are not interchangeable.

Current primary documentation shows why this matters:

• On the documented Adobe Firefly Video editor path, adding first or last frames disables Composition Reference, Motion reference, Shot size, Camera angle, and Style. Adobe's composition-reference clips have their own duration and size rules.
• Google documents first/last-frame and subject-reference flows separately, with availability tied to a specific Veo model and version.
• Runway may use several images to create a new still, then pass one resulting still to a video model. The image-stage allowance is not automatically the video-stage allowance.
• MiniMax documents model-specific totals for images, videos, audio, and combined files in H3 reference mode.

Genflow's live model selector is the source of truth for what Genflow currently exposes. Recheck it on production day. A base model may support an input that the current Genflow integration does not yet surface. Never convert a vendor's backend limit into a Genflow promise without verifying the active path.

Record both per-file limits and aggregate limits. “Each clip is short enough” does not prove that the sum of all video references is accepted.

E — Evidence: attach provenance and permission

For every asset, preserve its origin and permission trail. At minimum, record creator or supplier, acquisition date, allowed use, territory and term, restrictions, and the reviewer who approved it. For a recognizable person, record the likeness or performance authorization required by your organization.

Standards such as IPTC Video Metadata Hub and C2PA offer richer ways to express metadata and provenance. Use them when your workflow supports them, but keep the distinction clear: provenance evidence can show history; it does not by itself prove that a use is lawful, truthful, or within the license.

This checklist is a production record, not legal advice. Escalate ambiguous rights, minors, sensitive likenesses, trademarks, and restricted client material to the responsible legal or policy owner before upload.

C — Compatibility: catch mutually exclusive controls

Make a small compatibility matrix for the selected path:

Requested control	Accepted in this mode?	Consumes which budget?	Conflicts with
First frame		Image/keyframe
Last frame		Image/keyframe
Subject/product reference		Image/reference
Composition or action video		Video/reference
Performance driver		Video/specialized mode
Voice or audio reference		Audio/reference
Source video to modify		Video/edit mode

Fill the matrix from the current Genflow controls and linked vendor documentation. If a combination is unsupported, split it into stages. A common route is to use several references to create an approved still, then use that still as the sole image input for the video step.

Do not assume that controls commute. A keyframe plus a motion reference can mean “two useful constraints” in your brief while being an impossible combination in the selected interface.

A — Asset health: inspect the material, not just the extension

For images, inspect the delivery crop and full resolution. Reject blurred faces, malformed hands touching the product, illegible labels, incorrect variants, baked-in overlays, inconsistent reflections, and accidental background objects. Runway's current image-to-video guidance warns that visible defects in the input can be intensified during animation.

For videos, record the usable in/out range, duration, resolution, frame rate, codec, camera motion, cuts, occlusions, and whether the intended action is fully visible. A reference with three edits may be a poor motion source even when its file format passes.

For audio, listen for intelligibility, background noise, clipping, abrupt edits, language, speaker identity, and material that the model should ignore. Do not invent a universal loudness threshold: upload requirements and creative tolerances differ by tool and delivery channel.

Tools such as ffprobe can automate container, stream, codec, and duration capture. Use a separate whole-file SHA-256 tool for the Asset Contract's stable checksum; ffprobe packet or codec-extradata hashes do not identify the complete source file. A crop, trim, or transcode saved for upload is a new version and needs its own whole-file checksum. Neither inspection nor hashing can determine whether the correct product appears, whether the performance fits the brief, or whether you have permission to use it.

P — Pilot: begin with the smallest viable pack

Plan the first run around one approved subject or product master and one necessary control asset. Add audio only when the selected mode supports it and the brief requires it. Predefine which observable requirement would justify adding each optional asset; route actual failed-output diagnosis to the prompt troubleshooting log.

This is not a claim that fewer references always produce better video. It is a planning discipline: a smaller starting pack makes ownership and conflicts visible. Optional sources earn a place only when they own a requirement the initial pack cannot express.

The pre-run release gate

Before the first paid run, the producer and asset owner should be able to answer yes to all of these:

• Every asset has one primary role and a stable ID.
• Each fixed fact has exactly one approved master.
• The selected Genflow model and mode accept each requested input role.
• Per-file and combined budgets pass for count, size, format, and duration.
• First/last frames and other reference controls are not mutually exclusive in this path.
• Images, video, and audio pass technical and visible-content inspection.
• Source, license, likeness consent, restrictions, and approver are recorded.
• Contradictory versions and unnecessary references are removed.
• The first run uses the smallest pack that can satisfy the brief.
• The generation receipt will preserve model, mode, settings, asset IDs, and date.

If one answer is no, hold the run. A clear hold is cheaper to diagnose than a generation whose inputs cannot agree.

For example, a product master may be included as “product truth,” a five-second camera clip included as “composition,” and a mood board excluded because its older packaging conflicts with the master. If the chosen mode does not accept composition video with a first frame, the contract must choose one route or split the workflow into stages; it cannot mark both controls as accepted.

How this maps to Genflow

Genflow brings multiple image and video models into reusable workflows, but controls vary by provider, model, version, and mode. Use the image-to-video tool for an approved still, the text-to-video tool when visual reference is unnecessary, and the current Studio model controls to verify multimodal inputs.

ROLE-CAP and the Asset Contract are Genflow Editorial worksheets. They do not claim that Genflow automatically checks rights, resolves conflicts, or validates every vendor limit. Your team still owns the brief, authorized source material, current capability check, and release decision.

Research and preparation notes

Genflow Editorial reviewed 17 records from official model documentation, vendor workflow documentation, research, standards, and current Genflow product source. Automated tools helped organize the evidence, draft the article, check overlap, and create the conceptual cover. An independent reviewer evaluated the article before publication.

The cover is an original metaphor for an asset-routing tray. It is not a Genflow interface, customer content, model output comparison, or proof of successful generation. No model quality, success rate, savings, or rights outcome is claimed.

The pack is ready when every input has an owner

A reference pack is not ready because every available file has been attached. It is ready when each included file has a distinct job, a compatible route, healthy media, recorded permission, and no unresolved conflict with the approved truth. That decision record gives the first generation a testable contract—and gives the team a clean place to look when the result misses it.