Quality Assurance

U-Gen uses AI vision models to evaluate every generated keyframe and video segment before proceeding to the next pipeline stage. If content doesn't meet the quality threshold, it's automatically regenerated.

QA models

Image QA (keyframes)

Primary: GPT-4o-mini

Fallback: Gemini 2.5 Flash

Product visibility, persona consistency, prohibited content (phones), interaction mode, composition.

Video QA (segments)

Primary: Gemini 2.5 Flash

Fallback: Gemini 3 Flash Preview

Motion quality, visual consistency, artifacts, product positioning throughout the segment.

Strictness modes

Choose a QA mode when creating a job to control how strictly content is evaluated:

Mode	Threshold	Best for
Lenient	0.55	Fast drafts, testing, high volume. Default for new jobs.
Standard	0.70	Most production use cases.
Strict	0.80	Premium content, brand campaigns.

Retries are configured separately with sliders: keyframe retries (0–5, default 1) and video retries (0–5, default 0). When QA is enabled, keyframe retries minimum is 1.

Default Mode

New jobs default to Lenient(0.55 threshold). Retry defaults are 1 keyframe retry and 0 video retries, but you can adjust both with sliders (0–5 each) in the wizard or pipeline settings.

Fail categories

When QA detects an issue, it's classified into one of these categories:

Instant fail

PHONE_DETECTED

Phone, camera, or recording device visible in frame.

Instant fail

MODE_VIOLATION

Interaction mode not respected (e.g., product not held when holdable).

High

PRODUCT_INTEGRITY

Product missing, distorted, or incorrectly represented.

High

CRITICAL_DEFECT

Major visual defect that makes the content unusable.

Medium

MOTION_ARTIFACT

Unnatural motion, jitter, or warping in video segments.

Medium

TEMPORAL_INCONSISTENCY

Visual inconsistency across frames (flickering, identity shift).

Low

VISUAL_ARTIFACT

Minor rendering artifacts (slight blur, compression noise).

Low

BELOW_THRESHOLD

Overall quality score fell below the configured threshold.

Phone Detection

Phones, cameras, and recording devices in frame trigger an instant QA failure regardless of the quality score. This is a hard rule that cannot be overridden by any strictness mode.

Failure notifications

If you configure a Telegram chat ID in your settings, the system sends a notification via Telegram when a job fails QA after exhausting all retries. The message includes failure details and job ID. This is optional and disabled by default.