Defect Generation

Synthesize realistic, localized industrial defects onto defect-free (OK) part images with automatic COCO bounding boxes and segmentation masks.

​

Overview

Defect Generation on the Repli5 platform synthesizes realistic physical defects (scratches, dents, tears, stains, weld flaws, wrinkles) directly onto defect-free (OK) product photos.

For every generated image, the Repli5 platform automatically computes the pixel-level difference against the untouched OK source image, blends the defect seamlessly into the part’s native surface grain, and exports COCO segmentation masks and 2D bounding boxes (annotations_coco.json).

---

​

Key Concepts: Presets & Templates

• Preset: Defines one defect type—its COCO class name, visual description and placement, optional reference close-up photos, and paste/blend settings. You can configure up to 8 presets per template in Separate mode (up to 5 in Combined mode).
• Template: Saves a collection of presets so you and your team can reload the exact same defect suite across future datasets. Clicking Load template lists all templates saved to the active project as well as your personal templates.

---

​

Step-by-Step Guide

1. 1. Defect-Free (OK) Images — Select Source Photos

   In Card 1, choose your defect-free (OK) base images:

   • Dataset tab (Default): Pick an uploaded dataset of OK part photos from your library or active project, and select specific images on the page or the entire dataset.
   • Upload tab: Upload a single OK part image from your computer for quick testing and preset tuning.
2. 2. Defect Templates & Presets — Configure Defects

   In Card 2 (Defect templates), load an existing template from the Template Bar or create presets from scratch (click + New, Rename, Duplicate, or Delete):

   • Defect Class: A short label (up to 40 characters) used as the exported COCO category name (e.g., scratch, dent, wrinkle, tear). Presets that share the same Defect Class name automatically share the same COCO category_id across the dataset.
   • Defect Description & Placement: One unified text box (with no character limit) describing what the defect looks like (type, shape, depth, edges, color, contrast), how big it is (e.g., fraction of the part or cm/mm), and where to place it (using visible part landmarks). If you omit a location, the AI planner automatically picks a suitable spot on the part at random for each image.
   • Enhance Button: Click Enhance below the description box to have the AI expand a brief note into a detailed, production-ready defect description tailored to your Defect Class and reference images.
   • Randomized Choice Groups: Use flat curly-brace choices such as a {shallow hairline scratch | deep gouge}, about {2% | 5%} of the part's length to pick one option at random for each generated image.
   • Reference Images (Optional, up to 3): Upload up to 3 close-up photos of a real defect (PNG, JPEG, or WEBP). Only the defect’s shape and texture are copied—never its background color, lighting, or position. When reference images are attached, the text description becomes optional.
   • Paste & Blend Controls:
     • Faithful paste (Default ON): Pastes only the ΔE-gated, low-pass change so the original photo’s sensor grain and micro-texture show through. Turn OFF to paste the generated defect region directly (useful when moving existing seams or stitching).
     • Blend (Optional): When unchecked, the defect is applied at full strength (100%). Check Blend to specify a custom strength range between 10% and 100% (e.g., 50 to 70 %), drawn randomly per image.
3. 3. Generation Mode & Advanced Options

   In Card 3 (Summary), configure how multiple presets are applied:

   • Separate Mode (Default): Outputs 1 image per preset for each selected OK image (N images × P presets = N × P tasks, up to 8 presets). Best for building balanced single-defect training sets.
   • Combined Mode: Applies all presets onto each OK image (N images = N tasks, up to 5 presets). Best for multi-defect inspection scenarios.
   • Negative Decoy Paste (Optional, Default OFF): Pastes one unlabeled defect-free patch on a non-overlapping spot of the surface using the same blend edge. Keep OFF by default when training with OK images or on low-contrast defects; enable only when training without OK images to suppress paste-boundary false positives.
4. 4. Summary, Credits & Launch

   In Card 3 (Summary):

   1. Review the total task count and Required credits (1 credit per task).
   2. Verify the Evaluation Rubric Set (defaults to Defect generation, which scores Defect Realism, Surface Continuity, and Scale & Geometry).
   3. Click Augment to start generation.

---

​

Writing Effective Defect Prompts

Do (Recommended)	Avoid
”a shallow linear scratch with bright bare metal and jagged edges, ≈3% of the part’s longer side, on the top-left corner of the bracket"	"scratch” (no shape, contrast, or size information)
“a soft but clearly visible dent with a distinct bright highlight next to a dark shadow, borders fade out, no sharp rim"	"a subtle low-contrast mark somewhere on the part” (too faint to pass diff masking)
“an unbroken dark scuff line, equally dark from end to end, along the outer bezel edge”	Nested choice groups (use flat single-level choice groups only)

---

​

Output Dataset & COCO Export Structure

When a Defect Generation job completes, its output dataset in the Data Library includes:

• rgb_out/: The synthesized defective part images (NOK).
• rgb_raw/: The original defect-free source images (OK).
• instances/ & semantics_rgb/: Pixel-accurate defect masks derived from the OK-vs-NOK difference.
• annotations_coco.json: Ready-to-train COCO .json annotations containing images, categories (mapped from your Defect Class names), and annotations (bbox, segmentation polygons, and area).