Define the sources of consistency risks before project initiation.
In generative brand visuals, character and product consistency primarily concerns continuity risks throughout the production workflow. During project initiation, brands must define consistency across six dimensions: character appearance, product form, brand color palette, scene logic, motion rhythm, and audio tone. Missing any element will be perceived by viewers as disjointed in the final cut.
Specifically, have the brand provide a consistency conflict checklist detailing past issues such as face-swapping errors, product distortion, color drift, or proportion mismatches. The production team uses this list to identify error-prone stages and decide whether to use fixed character references, locked 3D product models, or frame-by-frame post-production verification. The criterion is simple: if the brand cannot provide at least three historical failure cases, project complexity is low and workflows can be simplified; if failures concentrate on character faces, a dedicated character identity library must be established.
The risk is that brands often equate consistency solely with unchanged character appearance, overlooking product details and scene lighting. An exception applies when projects pursue stylized expression, such as abstract ads or concept shorts; consistency requirements may be relaxed but must be confirmed in writing beforehand. The consequence is that failing to define consistency parameters during project initiation will multiply rework costs and may even prevent on-time delivery.
More reference materials from the brand do not necessarily yield better results.
Many brands assume that providing numerous reference images guarantees consistency, but the opposite is often true. Piling up materials confuses generative models, exacerbating character and product drift. Effective preparation involves providing three to five high-resolution images per core character and product, showing varied angles, lighting, and expressions, accompanied by text specifying which features must remain fixed and which may vary.
The production team must verify that the brand holds copyrights for these assets, especially portrait rights and product design patents. If real actors are used, portrait right agreements must be signed, clearly defining the scope of use for AI-generated versions. For products with unique textures or packaging, original design files are required rather than photographs of photos. The standard is that all reference materials must be traceable to their original sources and have a resolution no less than 70% of the final delivery specification.
The risk is that filters or retouching in brand-provided reference images cause generated results to deviate from the actual product. An exception applies when projects require exaggerated or anthropomorphized products; stylized references are permissible provided allowable distortions are clearly marked. The consequence is that incomplete materials prolong early testing cycles, potentially leading brands to mistakenly attribute delays to team inefficiency rather than insufficient input quality.
Storyboards must lock keyframes for characters and products.
Storyboards for generative commercials cannot list only actions and dialogue; they must specify keyframe requirements for characters and products in every shot. Keyframes refer to quantifiable parameters such as facial angle, product orientation, light direction, and color range. During storyboarding, the production team must define these parameters via text or simple diagrams rather than relying on random AI generation.
Specifically, the production team breaks down the storyboard into a shot-level checklist, where each shot includes character ID, product ID, camera movement, shot size, duration, and consistency checkpoints. The brand must confirm each checkpoint individually, particularly verifying logo clarity and costume continuity. The criterion is that if a shot's consistency requirements cannot be clearly described in text, it is unsuitable for generative production and should switch to live-action or CGI.
The risk is that over-reliance on AI generation capabilities in storyboarding makes keyframes irreproducible. An exception applies when projects allow different costumes across scenes; in such cases, only facial features need locking, while product appearance must remain consistent throughout. The consequence is that keyframes undefined during storyboarding are nearly impossible to fix later, forcing reshoots or regeneration and causing budget and schedule overruns.
How to Allow Margin for Generative Post-Production During Shooting
Even when primary visuals are AI-generated, the filming phase must still capture clean plates for post-production compositing. Brands and production teams must jointly decide which elements require live-action shooting and which can be replaced with generated imagery. Product close-ups, extreme facial close-ups, and scenes requiring realistic lighting interactions are better suited for live action, as AI generation often distorts these details.
Specifically, use gray and color charts during filming to record ambient lighting and provide a baseline for color grading. Simultaneously, capture empty background plates without actors as references for background replacement or AI generation. The production team must ensure that the resolution, frame rate, and color space of live-action footage match the specifications of generated assets; otherwise, edge flickering or color banding may occur during compositing. If the specification discrepancy between live-action and generated assets exceeds 10%, reshooting or regeneration is mandatory.
A key risk is brands reducing live-action shots to cut costs, resulting in insufficient reliable anchor points for compositing. An exception applies to fully virtual projects requiring no live action, provided 3D models or depth maps are supplied as generative references. Consequently, insufficient footage during filming forces the post-production team into repetitive generation attempts, potentially yielding a final video with obvious AI artifacts that the brand cannot accept.
Post-production must establish frame-by-frame verification and version control mechanisms.
Post-production for generative video focuses on multiple iterations. The production team must implement a frame-by-frame verification process, checking character facial features, product details, color consistency, and motion logic for every shot. Brands should participate in milestone approvals rather than reviewing only the final cut. Three approval milestones are recommended: rough cut, fine cut, and final color grading.
Specifically, the production team should issue a consistency report at each milestone, listing approved shots, required revisions, reasons, and proposed solutions. Brands must provide itemized feedback on the report, avoiding vague comments like "it feels wrong" by specifying issues such as character eyes, product logos, or background colors. If a single shot fails approval after three consecutive revisions, the team must revert to storyboarding or filming for re-evaluation instead of continuing to deplete the post-production budget.
A risk arises when brands frequently introduce new creative requests during post-production, causing constant shifts in consistency baselines. For projects with strict release schedules, an exception allows setting a revision cap, permitting only minor adjustments thereafter. Without version control, the final deliverable may mix multiple versions with disorganized source files, making future modifications untraceable.
Acceptance checklists must cover both technical metrics and brand standards.
Final video acceptance requires more than visual appeal; it demands item-by-item verification against consistency dimensions defined during project initiation. Brands should prepare an acceptance checklist covering six areas: character consistency, product consistency, color consistency, audio consistency, subtitle accuracy, and copyright compliance. Each section requires clear pass criteria, such as character facial features deviating no more than 5% from reference images and product logo clarity matching the original design files.
Specifically, during acceptance, brands should compare frame-by-frame screenshots of the final video side-by-side with reference images to document discrepancies. The production team must provide source files, project files, font files, and music licenses to enable the brand's future edits or redistribution. If any acceptance criterion fails, the production team must rectify the issue free of charge within the agreed timeframe and resubmit for approval.
The risk is that brands focus solely on visuals while neglecting audio and subtitle consistency, causing sync issues or misaligned captions across platforms. An exception applies to social media short videos, where audio acceptance can be simplified, provided auto-generated captions do not obscure key information. The consequence of delivery is an incomplete acceptance checklist; the brand discovers issues post-launch, but the production team has already completed delivery and cannot be held accountable.
Applicable boundaries and exclusion criteria must be clarified in advance.
AIGC commercials and AI product videos are not suitable for all brand projects. If a product features highly precise physical structures, such as medical devices or car engines, generative AI may fail to accurately render internal details, making live-action or 3D animation preferable. If strict legal compliance is required, such as for pharmaceuticals or financial products, AI-generated content may not pass review and will require additional manual verification.
Specifically, brands should use an applicability self-assessment form before project initiation to evaluate suitability for generative production. This form covers five dimensions: product complexity, brand guideline strictness, platform requirements, budget, and schedule flexibility. The production team must clearly advise against AIGC if the project falls outside applicable scope and provide alternative solutions. The criterion is that if more than three items on the self-assessment are high-risk, the project is unsuitable for AIGC.
The risk is that production teams conceal inapplicability to secure contracts, leading to mid-project failure. An exception exists when the brand has a dedicated AI team capable of handling generated outputs independently, allowing the production company to handle only specific stages using partial AIGC. The consequence is that failing to define boundaries upfront leads to substandard results after significant resource investment, damaging relationships and industry reputation.
The next step is to start with a small-scale test.
Brands attempting AIGC commercials for the first time should avoid launching full projects immediately. We recommend starting with a two-week test featuring a single product, character, and scene. The primary goal is to verify control over character and product consistency and assess the team's problem-solving capabilities. Expanding to a full project only after passing this test significantly reduces risk.
During the testing phase, the brand provides the aforementioned materials, and the production team delivers three test clips in different styles along with a consistency report. The brand uses this report to decide whether to proceed and identify necessary adjustments. This test costs far less than a full project yet exposes most potential issues. If results are unsatisfactory, the brand can cut losses promptly and avoid further investment.
If you are preparing an AIGC commercial project, please gather your brief, visual references, product or corporate materials, delivery platforms, and copyright scope before viewing ourAIGC Video Services pageTranslate abstract preferences into actionable production boundaries through communication.
