A Quality-Control Checklist Workflow

AI image and video generation is probabilistic — the same prompt and settings will occasionally produce assets with defects that are statistically unlikely but not impossible. Common defects include anatomical errors in human subjects (fused fingers, extra limbs, asymmetric facial features), text artifacts (garbled letters, inconsistent font rendering, text overrunning its box), composition errors (subject partially clipped at frame edge, excessive background headroom, unintended object intrusions), and technical artifacts (JPEG blocking at low quality settings, color banding in gradients, noise patterns in flat areas, soft focus on intended sharp subjects).

When reviewing 10 assets manually, these defects are easy to spot. When reviewing 200 assets from a batch run, fatigue and speed cause subtle defects to slip through. Research consistently shows that human reviewers miss 15–30% of subtle defects in large-batch visual review sessions. For campaigns where a defective asset could go live in a paid ad or a client presentation, this miss rate is unacceptable.

Automated QC addresses this by running deterministic analysis nodes that do not fatigue, do not vary with reviewer mood, and produce the same score for the same defect every time. The workflow runs faster than human review and produces a structured log that can be audited, filtered, and reported on — enabling trend analysis such as "this model produces hand defects at 8% rate on prompts with this pattern" which feeds back into prompt engineering improvement over time.

The QC rubric is a configurable set of checks, each with a pass/fail threshold and a severity weight. A typical production rubric includes checks in four categories: anatomical correctness, composition validity, technical quality, and brand compliance. Each check is an analysis node that takes the asset as input and returns a score between 0 and 1 (or a categorical result such as "pass/flag/fail"). The rubric aggregates these scores into an overall QC result using a weighted formula.

Anatomy checks: face detection node confirms a face is present (for portrait assets) and returns a face quality score (symmetry, eye openness, absence of artifacts); hand detection node confirms hands are present at the expected count with correct finger count (five per hand detected). Composition checks: edge clipping detector confirms the primary subject does not overlap the frame edge within a configurable margin; headroom calculator confirms the composition has appropriate space above the subject; intrusion detector flags unexpected objects in the scene that do not match the prompt description.

Technical quality checks: resolution validator confirms the output meets the minimum resolution requirement; sharpness scorer measures local contrast in critical regions (faces, product labels, text); color banding detector analyzes gradient smoothness in background regions; noise level scorer measures signal-to-noise ratio in flat areas. Brand compliance checks: color temperature scorer compares the asset's dominant color temperature to the brand's approved range; logo presence checker confirms the watermark or logo is present and correctly positioned (for assets processed by the branding workflow).

Store the rubric as a named configuration in the QC Config node. Different use cases may have different rubrics: a product photography QC rubric focuses on product clarity and background cleanliness; a social media content rubric focuses on composition and color temperature; a character illustration rubric focuses on anatomy and style consistency.

When the QC rubric flags an asset as failing, the workflow routes it to a correction branch rather than simply rejecting it. The correction branch applies targeted remediation based on the specific failure reason: an anatomy failure routes to a face/hand repair node; a composition failure routes to a reframe-and-crop node; a sharpness failure routes to an upscale-and-sharpen node; a color temperature failure routes to a color correction node.

The correction nodes are not full regenerations — they are targeted image-processing or low-denoising-strength refinement passes that fix the specific defect without altering the parts of the asset that already passed QC. This distinction matters for credit efficiency: a targeted sharpness correction costs a fraction of a full re-generation, and for assets that only fail on one or two checks, targeted correction achieves the same result as re-generation at much lower cost.

After correction, the asset re-enters the QC rubric for re-scoring. If it passes on the second evaluation, it proceeds to the output queue as "passed with correction" — flagged in the log as having required remediation so the production team can review corrected assets for sanity-check even though they passed QC. If it fails the second evaluation, it is routed to the "manual review required" queue where a human reviewer makes the final decision. This three-tier system — auto-pass, auto-correct, manual review — handles the vast majority of assets automatically while surfacing only the genuinely ambiguous cases for human attention.

Every asset processed by the QC workflow produces a structured QC record containing: the asset ID, the run timestamp, the result for each individual check (with numeric score), the overall QC result (pass/fail/corrected), the correction type applied (if any), and the final disposition (auto-passed, auto-corrected, sent to manual review). These records are aggregated into a QC report for each batch run.

The QC report is the primary interface between the automated QC system and the production team. Configure the Report Generator node to produce the report in your preferred format: a JSON summary suitable for ingestion into a project management or DAM system, a CSV suitable for spreadsheet review, or an HTML summary with thumbnail previews of each asset and its scores. The HTML report format is the most useful for visual review — it shows each asset alongside its QC scores, highlights the specific defect that triggered the failure in red, and shows the before/after comparison for corrected assets.

Aggregate the QC report data over time to identify systemic quality issues. If a particular model configuration consistently fails the hand anatomy check at above 10% rate, the prompt template for that model should be updated to include hand-specific quality tokens. If a specific output resolution consistently produces sharpness failures, the upscaling settings should be adjusted upstream. The QC report is not just a quality gate — it is a feedback loop that improves the generation quality of upstream nodes over time, progressively reducing the failure rate as prompt and model configurations are tuned based on observed QC data.

The QC checklist workflow is most powerful when it is integrated as a mandatory final stage in every upstream generation workflow rather than as a standalone post-processing step. In the Floniks editor, build QC integration by connecting the output of any generation workflow to the QC workflow's input via a Workflow Chaining node. This node automatically triggers the QC workflow when the upstream generation completes and passes the generated assets as the QC input batch.

With this integration, no generated asset can reach the output queue without passing the QC rubric. The production team's asset library only ever contains QC-passed or QC-corrected assets, eliminating the separate "review before use" step that manual QC requires. The only assets requiring human attention are those in the "manual review required" queue, which for a well-tuned rubric represents 2–5% of total output — manageable even at high batch volumes.

For client-facing deliverables, add a digital sign-off node at the end of the QC workflow that attaches a QC certificate to each delivered asset: a metadata annotation listing the QC checks performed, the scores achieved, and the "QC passed" timestamp. This certificate serves as documentation in the event of a post-delivery dispute about asset quality and demonstrates to clients that a systematic quality process was applied to every deliverable, not just those the team happened to spot-check.