Management space for failed samples must be reserved during the project initiation phase.
When brands launch AIGC commercial or AI product video projects, they usually only focus on the final video effect and rarely consider how to handle failed samples. However, in actual production, a massive amount of time is consumed by repeatedly generating, screening, and modifying samples; without a systematic screening and archiving method, both the project timeline and budget will spiral out of control.
During project initiation, the brand and the production team should jointly confirm three things. First, clearly define failed samples: do they refer to technically unqualified generated results, versions that deviate from the creative direction, or assets that cannot be used in post-production compositing? Second, designate the decision-maker for sample screening—whether it is the brand's marketing director, creative director, or the production team's technical lead—to avoid process chaos caused by multiple people giving feedback simultaneously. Third, establish the format and storage location for sample archiving; it is recommended to name files by date, shot number, and version number, and sync them to a shared cloud to ensure all relevant personnel can view them at any time.
The criterion is that if the process for handling failed footage is not discussed at project kickoff, the post-production phase will likely suffer from version confusion, high rework costs, and low decision-making efficiency. The risk is that the brand may delay the launch due to repeated revisions, while the production team may continuously repeat useless work due to a lack of clear feedback. An exception is that if the project cycle is extremely short and the budget is sufficient, the archiving step can be skipped to proceed directly to final video production, but this situation is rare in brand projects.
How to reserve failed assets for AI generation during the shooting phase
Brand AI video projects often combine live-action shooting with AIGC generation, so the shooting phase must consider which assets might be deemed failures and how to retain them for subsequent analysis. During live-action shooting, it is recommended to shoot several backup takes for each shot, especially for scenes involving character movements, expressions, and lighting changes, because these details are prone to distortion during subsequent AI generation processing.
Specifically, the camera crew should record the technical parameters for each shot on set, including aperture, shutter speed, ISO, white balance, and lens focal length, while also noting the shooting intent and potential failure risks. For example, moving shots are prone to motion blur, and low-light environments are prone to noise, both of which can become causes of failure in post-production AI processing. The production team should archive this information along with the dailies to facilitate troubleshooting in post-production.
The criterion is that if at least two to three different versions of each shot are retained in the shooting assets and the technical parameters are recorded, there will be sufficient base material for experimentation during post-production AI generation. The risk is that the shooting crew may ignore backup takes due to tight schedules, making it impossible to cover certain creative directions in post-production. An exception is that if the project relies entirely on pure AI generation without live-action shooting, the archiving work in the shooting phase can be omitted, but the generation parameters and seed values must be clearly recorded.
Establishing a classification and screening process for failed footage in the post-production phase
Post-production is the stage where the most failed footage appears, especially during the AIGC generation process, where model outputs often contain a large number of images that do not meet expectations. It is recommended to divide failed footage into three categories for separate processing: technical defects, creative deviations, and style inconsistencies.
Technical defects include issues such as image tearing, character deformation, unnatural lighting, and insufficient resolution; this type of footage usually cannot be resolved through simple retouching and requires regeneration or parameter adjustment. Creative deviations refer to image content that does not match the script description, such as generating a scene panorama when a product close-up is required; this type of footage requires re-entering prompts or modifying the storyboard. Style inconsistencies refer to generated images that do not match the brand tone, such as generating a retro result when the brand requires a minimalist style; this type of footage requires adjusting style reference images or model weights.
The screening process is recommended to be divided into three steps. In the first step, the post-production technical lead conducts an initial screening to eliminate footage with obvious technical defects and retain versions with modification potential. In the second step, the creative team conducts a secondary screening to select versions that align with the creative direction from the retained footage and mark specific areas that need modification. In the third step, the brand participates in the final review to confirm the footage to be used and clarify which failed footage needs to be archived for future reference.
The criterion is that if the number of failed footage retained for each shot does not exceed three, and each piece of footage has a clear record of the reason for failure, then the screening process is effective. The risk is that valuable footage may be overlooked due to subjective preferences during the screening process, leading to restricted creative directions. An exception is that if the project timeline is extremely tight, the initial and secondary screenings can be skipped entirely, allowing the brand to directly decide the final version, but this will reduce the quality of the final video.
Archiving standards determine the reuse value of failed samples.
Failed samples are not useless; they can serve as reference materials for future projects or be used to train brand-specific AI models. Therefore, archiving cannot be simple storage, but must follow structured standards.
When archiving, each failed sample should include the following information: generation time, models and parameters used, prompt text, failure reason tags, revision history, and a link to the corresponding successful sample. For example, a sample with blurred product edges must record the specific location of the blur, the repair methods attempted, and the parameter differences from the final successful version. This way, when similar issues arise later, historical records can be accessed directly to avoid repeated trial and error.
It is recommended to organize the storage structure into folders by project name, date, shot number, and version number, and to include key tags in the file names, such as "Technical Defect-Edge Blur-20250301-Shot 03-v2". Additionally, at the end of the project, it is recommended to compile the archive directory into a document containing an index and descriptions of all failed samples for the brand to store.
The standard for success is that if an archived failed sample can be found and its failure reason understood within ten minutes, the archiving is successful. The risk is that incomplete archiving information will prevent future reuse, or that the archive files will become too large and consume storage space. An exception is that if the project is a one-time release with no subsequent iterations, archiving can be simplified to retain only the final deliverables and a few key failed samples.
How to use failed samples during the acceptance phase to improve final video quality.
When accepting the final video, the brand should not only look at the final version but also evaluate the production team's work quality in conjunction with the failed samples. By comparing the failed samples with the final video, it is possible to determine whether the team iterated sufficiently, resolved key issues, and if there is still room for improvement.
Specific acceptance actions include requiring the production team to provide a list of failed samples for each shot and explaining the reason each sample was rejected. The brand can randomly spot-check a few shots, compare the differences between the failed samples and the final video, and confirm that the production team has not overlooked obvious issues. At the same time, the acceptance process should verify that the archive is complete, including failed samples, parameter records, and revision logs, as these materials form the foundation for future maintenance and optimization.
The standard is that if every shot in the final video can be traced back to at least one failed sample, and the failure reasons align with the final revision direction, the acceptance is qualified. The risk is that the production team might provide only a few failed samples to conceal issues, preventing the brand from conducting a comprehensive evaluation. An exception is that if the project uses extensive AI generation and the process is highly automated, the number of failed samples may be massive; in this case, sampling can be used for acceptance, but it must ensure coverage of all key shots.
Situations where failed sample management is not applicable.
Not all brand AI video projects require a strict process for screening and archiving failed takes. The following situations allow for simplifying or skipping this process.
The first situation is pure creative exploration projects, such as AI-generated videos used internally by the brand for brainstorming, which are not used for external distribution; in this case, takes can be generated and discarded at will without archiving. The second situation is rapid-response projects, such as trending content on social media that needs to be published within a few hours; in this case, time does not allow for systematic management, and the production team can only rely on experience to make quick decisions. The third situation is projects with extremely low budgets, where archiving requires additional time and labor costs; if the budget cannot cover this, it is recommended to keep only the final video and a few key assets.
The criteria for judgment are that if the project cycle is shorter than a week, the budget is below the standard level, or the final video is only for internal reference, a complete archiving process does not need to be established. The risk is that simplifying the process may make subsequent revisions difficult, but when weighing the options, rapid delivery may be more important. An exception is that even if the project cycle is short, if the brand expects to reuse the assets or create derivative works in the future, at least the generation parameters and seed values should be retained for future traceability.
Next Steps
When launching the next AI video project, the brand should consider holding a dedicated meeting with the production team to discuss the screening and archiving plan for failed takes. Clarify which shots need to be kept as backup takes, which parameters need to be recorded, and which reasons for failure need to be annotated. Even if the project is small, it is recommended to establish simple archiving habits, because the uncertainty of AI generation means that failed takes are the norm, and effectively managing these takes can significantly reduce project risks.
At the same time, you can pay attention to the publicly available service scope on the ONCE official website, including AIGC commercials, AI product videos, generative imagery, and brand AI video workflows, which can help brands establish more standardized production processes. However, please remember that any service needs to be evaluated in combination with specific project requirements, and templates should not be applied blindly.
If you are preparing an AIGC commercial project, you can first organize the brief, reference visuals, product or company materials, delivery platforms, and copyright scope, and then view theAIGC video service pageto ground the communication from abstract preferences to executable production boundaries.
