AI Image Upscaling for Print — Sagacitas Case Study

1. The Challenge

Generic upscalers fail at print art

A Direct-to-Film (DTF) print business paid a third-party API to upscale customer artwork to print resolution. Off-the-shelf upscalers are built for photos. On logos and lettering they blur thin strokes, invent detail and leave visible fringes around the design, and on printed film those flaws show.

The business problem: Better print quality, lower cost per image, and control over the technology — without changing how the rest of the production system worked. Any new solution had to plug in with zero production code changes.

2. How We Approached It

Model choice + print-aware engineering

The right model for print

Deployed Real-HAT-Sharper, a transformer-based super-resolution model whose sharp edge reconstruction suits lettering, logos and linework — the content that generic upscalers handle worst.

Print-aware engineering

Designed proprietary post-processing tuned to how DTF film actually prints, so output arrives print-ready with no manual clean-up.

Production-grade service

Packaged the work as a secured, API-first service on serverless GPUs. It scales to zero when idle, warms on demand, and plugs into the existing production system without code changes.

Measured optimisation

Profiled every stage, validated each change on real production artwork, and calibrated GPU settings for speed without giving up output accuracy.

3. Toolkit

Model Real-HAT-Sharper (4×)

Deep learning PyTorch Spandrel Hugging Face

Image processing OpenCV NumPy Pillow pyVips

Service & infra FastAPI Modal (serverless GPU) REST + Webhooks

4. Before / After Performance

Post-processing time (18.5 MP) −85%

before / after

1.94 s0.29 s

Data transfer overhead −87%

before / after

1,452 ms187 ms

5. Outcomes

Area	Before	After
Upscaling provider	Paid third-party API	Self-hosted Real-HAT-Sharper
Text & linework quality	Softened strokes, fringing	Crisp edges, print-ready output
Post-processing time	1.94 s	0.29 s
Dispatch timeouts	Recurring on large files	Eliminated
Infrastructure cost	Per-image fees	Pay-per-use GPU, scales to zero

Strategic Business Value

Vendor independence: A core production capability now runs on technology the business owns.
Better product: Customers' artwork prints sharper, with no manual clean-up required by staff.
Lower cost: No per-image fees and no GPU cost while idle — serverless scales to zero.
Low-risk rollout: Switching to the new service, or back, is a configuration change — zero production code changes required.

Note: Implementation details are proprietary to the client. Available for deeper technical discussion under NDA.