Research

Building the future of restorative AI.

We are a small lab working on one problem: recovering production-quality artwork from the files people actually have. These are working notes on the model, its evaluation, and the decisions behind it.

Model note

2026-08

v2 in production · internal eval

Alpha-native super-resolution for flat artwork

Print artwork is not a photograph: our upscaler reconstructs the alpha channel, the cut line of the print, natively, where general-purpose super-resolution returns flat RGB and the silhouette is merely stretched.

Off-the-shelf super-resolution models are trained on photographs. On flat artwork they hallucinate glyph texture (an e becomes an o), wash discrete design colours toward means, and ignore transparency entirely. On a DTF transfer the alpha edge is the physical cut. Generic tools optimise how an image looks on a screen; prepress needs what survives on fabric.

Method

The model is trained exclusively on flat artwork with a print-realistic damage model (compression artefacts, resampling loss, colour bleed) on pairs where the original is known. It learns to reverse damage rather than to invent detail.

It is deliberately non-generative: no adversarial objective, because plausible-looking text is exactly how wrong text gets invented. It runs on commodity CPU.

Evaluation

Internal production designs with held-out ground truth. We measure what print cares about: edge sharpness relative to the true original, silhouette fidelity (shape IoU), palette drift against the design's own colours, and stroke survival, instead of PSNR/LPIPS, which reward photographic plausibility rather than print fidelity.

Results (internal evaluation)

Damaged reference logo, edge sharpness vs true original: 83% → 98%
Held-out designs restored to print bar without external vendor (n=20): 20/20
Vendor-escalation rate on the free tier (same eval): 45% → 0%
Ink mass preserved vs true original: ±1%

Baseline: against classical resampling (Lanczos) the same reference measures 24% of true sharpness.

Known limits

Detail that no longer exists in the source cannot be honestly recovered: a sub-pixel dot on an i stays lost rather than invented.
Extremely low-resolution sources below roughly a quarter of print resolution are refused rather than faked.
Fine connected script remains the hardest class; the system flags it instead of hiding it.

Method

2026-08

in production on every repair receipt · internal eval

A model that flags its own doubts

Every enhancement is inspected by an independent instrument that detects where fine detail may have closed up, and its count tracks real damage at r = 0.98.

Standard quality metrics are blind to topology. A result can score near-perfect on sharpness and colour while the dot of an i has fused to its stem: one closed counter is a misprint the numbers never saw. The eye reads topology; most pipelines don't.

Method

The instrument compares the enhanced result against evidence still present in the damaged source and marks locations where ink was painted across what the source remembers as a gap.

It runs on every upscale and writes its verdict to the job receipt. Above a threshold, the customer is asked to review the marked spots rather than being shown silent perfection.

Results (10 designs with ground truth)

Real topology errors detected (recall): 89%
Detection count vs. true damage, correlation: r = 0.98
Per-detection precision: why it warns and never auto-edits: 0.28

We publish the precision on purpose: it explains why the instrument warns instead of cutting.

Known limits

A single detection is not proof, because legitimate script joins look similar. The signal is the count and the marked locations, reviewed by a human.
Repairing flagged spots automatically requires glyph-level understanding, which remains active research.

Evaluation

2026

in production · internal eval

Background removal built for artwork, not portraits

Segmentation models are trained on people and products. Artwork keying is a different problem (flat colour fields, deliberate holes, ink the same colour as the backdrop), and ours is built for exactly that, with the customer holding the final say.

Generic background removal decides by objectness learned from photos. Artwork breaks it: counters of letters must stay open, white ink on a white backdrop must survive, and a drop shadow may be content. A wrong call silently deletes part of a customer's design.

Method

A keying pipeline purpose-built for flat artwork, plus a learned corrector for the ambiguous regions.

Every uncertain region is surfaced to the customer as an explicit keep-or-remove choice, and every correction becomes a training example, so the system provably learns from its users.

Results (internal evaluation)

25-artwork benchmark against a leading general-purpose background-removal service: design better preserved: 24 / 25 (96%)
Silhouette overlap (IoU) vs. expert mattes: 0.91-1.00

Known limits

When the backdrop colour is also an ink colour, the information is genuinely destroyed; the system asks instead of guessing, and an opt-in expert pass exists for that class.

Cross-cutting

The generative tier is never autonomous

A generative reconstruction tier exists, but it only runs when the non-generative tiers cannot reach the print bar.

Its output must pass identity gates (the result's ink must be the original's ink, only larger) and it always lands in a human preview. Every tier attempt is written to a receipt.