Print-on-Demand QA
Print-specific metrics extending the core framework. Color fidelity for fabric printing, physical print constraints, transparency edge quality, input validation, and garment placement.
Designed for DTG and DTF print-on-demand operators running AI-assisted design pipelines. Catches the failures that look fine on screen and look terrible on the shirt.
1. Print-Ready Color & Ink Fidelity
AI designs in RGB may contain colours impossible to reproduce on a shirt. The on-screen preview will look vivid; the printed shirt will look dull or wrong.
CMYK gamut warning per printer profile
Why it matters
RGB designs may contain colours outside the printer's reproducible gamut.
Technical approach
Convert via the target printer's ICC profile; flag regions where Delta-E > threshold (e.g. > 6).
Brand colour out-of-gamut flag
Why it matters
Strict brand HEX values may fall outside the printer's CMYK gamut entirely.
Technical approach
Compare each approved HEX to its closest in-gamut CMYK reproduction; flag any exceeding client tolerance.
Total ink coverage check
Why it matters
DTG printers cannot deposit excessive ink. Coverage above 240–280% causes bleeding, long drying, and poor hand feel.
Technical approach
After CMYK conversion, sum C+M+Y+K per pixel; report the percentage of pixels exceeding the configured limit.
Rich black vs. pure black advisory
Why it matters
Pure black (0,0,0,100) can print as faded grey on some garments; rich black yields a deeper black.
Technical approach
Detect large solid areas with high K and low CMY; flag as advisory (informational, not pass/fail).
2. Resolution & Print Dimension Mapping
A 1024×1024 AI image may look sharp on screen but pixelate when printed at 12×12 inches.
Print size vs. image resolution check
Why it matters
Below the printer's minimum DPI, prints become visibly pixelated.
Technical approach
Effective DPI = pixels / intended print inches. Flag if below the printer minimum (e.g. 150 DPI for DTG, 300 for best quality).
Aspect ratio mismatch warning
Why it matters
Square designs forced into rectangular print areas get cropped or padded unexpectedly.
Technical approach
Compare image aspect ratio to expected print area; flag if deviation exceeds a small tolerance (e.g. 2%).
3. Transparency & Edge Quality
AI-generated cutouts often leave halos, fringes, or coloured spill at the subject edge — invisible on screen, ugly in print.
Fringing / halo detection
Why it matters
Halos of white or semi-transparent pixels print as a visible white border on coloured shirts.
Technical approach
Dilate the alpha edge mask. For pixels where 0 < alpha < 255, measure RGB distance to white (or known background). Flag dense edge regions matching background colour.
Semi-transparent edge pixel warning
Why it matters
Soft-edge transitions from AI upscaling or background removal print unpredictably.
Technical approach
Count pixels with alpha 1–254 in a narrow band (e.g. 2–3px) of the fully transparent region; flag continuous borders exceeding a threshold length.
Background colour spill
Why it matters
Removed backgrounds can bleed into the subject's edges as a coloured halo in print.
Technical approach
Compare chroma (Cb/Cr) of edge pixels against subject interior and removed background. Flag edges deviating toward background chroma.
Note · These three can be implemented as a single composite alpha-edge quality metric.
4. Pre-Generation Input Validation (Dimension 0)
New gate before generation. Failing any metric here blocks generation entirely — no point QC'ing output from bad input.
Uploaded sketch / reference image quality
Why it matters
Tiny, JPEG-compressed, or blurry input produces poor AI output regardless of model.
Technical approach
Check minimum resolution (e.g. 512px short side), BRISQUE, Laplacian variance for sharpness, and compression artifact detection.
Content moderation (copyright, NSFW)
Why it matters
Copyrighted logos or inappropriate imagery as input creates legal liability if printed.
Technical approach
Perceptual hash comparison against trademarked logo databases (sports, brands) plus NSFW classifier on the upload.
Watermark / stock agency logo detection
Why it matters
Watermarks (Shutterstock, Getty) on input would appear on the final printed product.
Technical approach
OCR for common watermark text, template matching, and perceptual hash comparison against a watermark library.
Note · These checks happen before AI generation begins. They form Dimension 0 — Input Validation, gated as Gate 0.
5. Design Placement & Print Area Safety
Elements placed too close to the print edge get cut off or sewn into seams. Fine details may not survive DTG printing.
Print-safe zone compliance
Why it matters
Edges may be trimmed or absorbed into collar / side-seam stitching during garment production.
Technical approach
Per-garment safe-zone margins (e.g. 1" from collar, 2" from side seams). Overlay a template mask; flag any non-transparent pixels outside the safe zone.
Minimum detail size check
Why it matters
Very fine lines or tiny text may be illegible or break apart in DTG printing.
Technical approach
Measure stroke widths via morphological skeletonization and distance transform. Flag stroke widths below a configurable minimum (e.g. 1–2mm at actual print size).
Integration into the framework
The addendum is incremental, not a rebuild. Of the 14 POD metrics, 6 absorb cleanly into 3 existing dimensions (D01, D02, D08); the remaining 8 require 2 new dimensions — D00 for pre-generation input validation, D10 for print-physics-only checks.
Dimension 0 — Input Validation (new)
New gate before current Gate 1. Owns metrics 4.1, 4.2, 4.3. Failing any metric here blocks generation entirely — no point QC’ing output from bad input.
Dimension 1 — Technical Delivery (extended)
Owns metrics 2.1 and 2.2 as a print_specs sub-dimension, alongside the ICC profile path and file-format suitability check (PNG with alpha for transparent designs).
Dimension 2 — Spatial & Texture Integrity (extended)
Owns metrics 3.1, 3.2, 3.3 as a single composite alpha-edge quality metric. These are general image-space artifacts; the same checks apply to any cutout deliverable, not just print.
Dimension 8 — Brand & Client Compliance (extended)
Owns metric 1.2 (out-of-gamut brand colours) as a client-specific check, alongside existing brand HEX Delta-E scoring.
Dimension 10 — Print Fidelity (new)
Owns metrics 1.1, 1.3, 1.4, 5.1, 5.2 — the print-physics-only checks (CMYK gamut, ink coverage, rich black, garment safe-zone, minimum detail size). These are meaningless on screen-only deliverables, which is why they earn a dedicated dimension rather than living inside D02.
Client Configuration Table
New fields: printer_icc_profile, min_dpi, max_ink_coverage, garment_safe_zones, min_stroke_width_mm, rich_black_preference.
Revised gating sequence
- Gate 0 — Input Validation (new)
- Gate 1 — Technical + Print Specs (DPI, format, aspect ratio)
- Gate 2 — Spatial Quality + Print Fidelity (artifacts, transparency, ink)
- … existing gates unchanged.
POD pilot engagements
We're scoping a small number of pilot engagements with print-on- demand operators running AI-assisted design pipelines. Pilots focus on Dimension 0 input validation and the alpha-edge quality composite — the two highest-leverage POD checks.
Get in touch