Skip to main content

​
Overview

The Variation Generator takes close-up RGB images of an object (either a single uploaded image or a dataset) and transforms their appearance—such as surface color, material finish, or ambient lighting—based on a natural-language prompt.
Defects & Geometry Are Preserved: Using edge-guided conditioning, the model preserves the exact object geometry as well as existing dents, corner wear, scratches, and surface defects while applying your requested variation.

​
Step-by-Step Guide

1

1. Upload / Dataset — Choose Your Input

In the Upload / Dataset card, choose whether to run on a saved dataset or a single uploaded photo:
  • Dataset tab: Select a dataset from your library or active project. You can select individual images on the page, Select all on page, or run across the entire dataset.
  • Upload tab: Click Upload image to upload a single close-up RGB image directly from your computer.
2

2. Variation — Describe the Desired Change

In the Variation text box, describe the visual change you want applied to the object or background (up to 2,000 characters).Good examples:
  • “Change the color of the phone housing to matte light gray”
  • “Change the metal bracket finish to brushed anodized black”
  • “Change the studio lighting to warm directional side-lighting”
You do not need to describe the existing scratches or dents in your prompt—the pipeline automatically locks the structural edges and surface wear from the source image.
3

3. Summary, Evaluation Rubric & Launch

In the Summary card:
  1. Verify your Input count and Variation prompt.
  2. Check the Required credits (1 credit per generated image) against your available balance.
  3. (Optional) Select an Evaluation Rubric Set (defaults to the Variation Generator quality rubric).
  4. Click Augment to enqueue the job.

​
Reviewing, Evaluating & Regenerating Results

Once submitted, your job appears in the Augment History panel at the bottom of the page:
  • Live Progress & Quality Scores: As tasks complete, an automated visual evaluation scores each output image against the selected rubric (1–5 scale). Images with any metric below 3 are flagged for review.
  • Inspect & Compare: Click any image tile to compare the original rgb_raw input against the generated rgb_out variation side-by-side.
  • Approve / Reject: Mark individual images as Approved or Rejected during quality review.
  • Regenerate Single Image: If an image has an artifact, click Regenerate on that tile (optionally adjusting the prompt) to replace it in-place without re-running the entire batch.
  • Export / Download: Completed batch runs automatically create an output dataset in your Data Library, ready for one-click ZIP download.