WORKFLOW • Framework updated · Jul 2026
AI Video Quality Control: A Scored Rubric Before You Spend More Credits
AI video tools bill you for the attempt, not the result. A discarded clip costs the same credits as a keeper, and none of the major billing pages cited here document a refund for an unusable generation. This guide gives you a scored approval rubric to apply before you buy another render, so the decision to regenerate is deliberate rather than reflexive.
Why a bad clip still costs full price
Every major platform meters generation on output, not outcome. Runway's developer pricing documentation bills video generation per second of output, with Gen-4.5 at 12 credits per second and Gen-4 Turbo at 5 credits per second, whether or not you keep the clip. Google Flow charges per generation rather than per request, with Veo 3.1 Lite at 10 credits, Veo 3.1 Fast at 20 credits, and Veo 3.1 Quality at 100 credits for each run. Luma prices by model, resolution, and duration. Across the Runway, Luma, and Google Flow billing pages cited here, none describe a refund or credit-back for a failed or unusable output. The practical consequence is simple: a second attempt is a second full charge. That makes a short, disciplined review of the first clip the cheapest quality-control step available to you.
The four-criteria approval rubric
Score each finished clip on four criteria before deciding to spend again: brief match, artifact check, brand safety, and technical spec. Rate each one pass, hold, or fail. A clip that passes all four is done, so ship it. A clip that fails any criterion goes back only if you can name the specific fix, because re-running the same prompt at the same settings tends to reproduce the same weakness at full cost. Treat hold as a prompt-language problem to solve at low or no charge, not a reason to buy a render. The rubric's job is to separate a clip that needs a different instruction from a clip that needs another expensive attempt. Gate the regeneration on the score, never on how many credits remain in your balance, since a leftover balance is not evidence that a clip is worth redoing.
Brief match: score every element the prompt promised
Google DeepMind's Veo prompt guide defines seven elements you can turn into a checklist: shot framing and motion, style, lighting, character descriptions, location, action, and dialogue. Read your original prompt, then score the delivered clip against each element you actually specified. Did the camera move as described? Is the lighting the mood you asked for? Does the character match, and does the action complete? A brief-match failure usually means the instruction was missing or ambiguous, not that the model got unlucky, so the fix is sharper prompt language rather than an identical re-roll. Mark brief match as pass only when every element you stated is present. If two or more elements are absent, rewrite those lines first, confirm the wording is specific and detailed, and only then consider spending on a new generation.
Artifact check: fix the prompt before you re-roll
An artifact check looks for the defects that make a clip unusable: warped hands, flickering, morphing faces, broken text, or unstable motion. When you find one, resist an immediate re-roll and change one variable at a time so you can see what each edit does. Google's Veo prompt guide advises that the more detail you add, the more control you have over the final output, so describe the fix concretely instead of re-rolling the identical prompt. For image-to-video work, refine the motion and camera language rather than re-describing what is already in the frame. Test cheaply before committing to a full render: Luma image generation runs at 4 credits per image in batches of four, or 16 credits per batch, which is a low-cost way to lock a look, and Luma's Relaxed queue mode allows iterating without drawing down fast credits. Only escalate to a paid video generation once the defect has a named cause.
Brand safety: confirm the clip is publishable
Brand safety asks whether the clip is publishable under the provider's own rules before you invest more credits polishing it. Google's Generative AI Prohibited Use Policy, which governs its generative tools, tells users not to engage in dangerous or illegal activities, not to compromise the security of others' or Google's services, not to engage in sexually explicit, violent, hateful, or harmful activities, and not to engage in misinformation, misrepresentation, or misleading activities, a category that includes misrepresenting the provenance of generated content. Each provider maintains an analogous policy, so read the one that governs the tool you are paying for. A clip that would violate those terms is not worth refining, because the account risk outweighs any credit already spent. Score brand safety as pass only when the footage, and the prompt that produced it, clearly sit inside the provider's policy and your own brand guidelines for depiction, likeness, and claims.
Technical spec: match resolution and duration to the need
Technical spec is where retry cost compounds, because higher resolution and longer duration cost proportionally more on every platform. Runway bills per second of output, so a longer clip is a larger charge on each attempt. Luma prices rise with model, resolution, and duration, and Google Flow's tiers climb from 10 credits for Veo 3.1 Lite to 100 credits for Veo 3.1 Quality per generation. Score technical spec against the clip's real destination: a short social cut rarely needs the highest tier, and a proof of concept does not need final resolution. Render at draft spec while brief match, artifact check, and brand safety are still in play, and reserve the expensive final settings for a clip that has already passed the other three. Note that Google Flow monthly credits do not roll over, but an expiring balance is a reason to plan, not to buy redundant high-spec retries.
Editorial note: This framework is general information, not a vendor endorsement. Check the current pricing, terms, and data-handling details directly with the provider before buying.