Skip to main content

​
Overview

Quality Evaluation & Review is a built-in quality assurance feature available across all augmentation workflows on the Repli5 platform. Whenever an augmentation job finishes generating an image, an automated vision-language evaluator compares the generated output (rgb_out) against the source image (rgb_raw)—and, for localized edits, close-up crops of the modified region—using the Evaluation Rubric Set selected for that job.

​
Workflow-Specific Evaluation Rubrics

Because each workflow has different quality criteria, each workflow defaults to its own dedicated Evaluation Rubric Set (which you can also customize or switch in Card 3 — Summary before launching a job):

​
Scoring Scale & Attention Flags

  • 1–5 Scoring Scale: Each metric in the active rubric set is scored from 1 (severe artifact or missing requirement) to 5 (excellent fidelity), accompanied by a short written explanation from the evaluator.
  • Needs Attention (Score < 3): Any image where at least one metric scores below 3 is automatically flagged in the Augment History so reviewers can inspect problem candidates first instead of paging through entire batches manually.

​
Reviewing & Regenerating Images in Augment History

In the Augment History panel at the bottom of any workflow page:
  1. Filter by Score or Review Status: Filter the batch by Needs Attention (< 3), Unreviewed, Approved, or Rejected.
  2. Side-by-Side Comparison: Click any image to open the interactive comparator (with synchronized zoom/pan, split curtain, and rubric score breakdown).
  3. Approve or Reject: Mark individual images as Approved or Rejected, or override a metric score with a human verdict.
  4. In-Place Single-Image Regeneration:
    • If an image has an artifact or low rubric score, click Regenerate directly on that image.
    • You can re-run with the same preset/prompt or adjust the prompt for that specific image.
    • When the new image completes, it is automatically re-evaluated and replaces the previous image in the output dataset (superseded = true), keeping your dataset count and annotations clean.