Bearings
A reference-first image system for a veteran newsletter.
Bearings readers catch a wrong uniform or insignia at a glance. I built a reference-first, rule-based image system, with AI in the pipeline and a review gate before anything ships. It produced image sets for 8+ long-form articles, and by my estimate it costs about 84% less than commissioned photography.
I turn a hard quality problem into a repeatable system: references first, rules second, tooling third, review before shipping.
The audience is the quality control.
Bearings publishes long-form essays for veterans on military career transitions, VA claims, grief, and leadership.
Its readers spot a wrong uniform or insignia at a glance. Stock imagery is generic and often wrong. Commissioned photography does not fit a small publication’s cadence. Typical AI imagery looks rendered and gets details wrong.
I needed images that looked photographed, got the details right, and could ship on the publication’s schedule.
Three things could not break.
- An audience that acts as quality control.
- One real founder’s likeness had to stay consistent across unrelated stories.
- Some essays concern real named people, who must never be rendered as likenesses.
See. Solve. Scale.
See
I gathered real references first: uniforms, gear, and real photos of the subject.
Solve
I designed each section’s images as a tonal arc of three. Cool and cramped at the opening, warm and open at the close.
Scale
I turned the rules into reusable tooling, including an HTML prompt builder.
The pipeline runs Midjourney for mood, Magnific for upscaling, and Google Nano Banana for grain and light. A review gate runs before anything ships. A failed review means I rebuild the prompt. I never patch the image.
Five decisions, five tradeoffs.
Start from real references, not generated ones.
Real reference photos held likeness. AI-generated references made it drift.
TradeoffSetup takes longer for each new subject.
Rebuild failed images. Never patch them.
The fix lives in the prompt, so the next article inherits it.
TradeoffMore regeneration time on each failure.
Show veteran identity through small artifacts.
A lapel pin or a dog-tag chain carries it. Flags do not.
TradeoffThe images are quieter and less instantly “military.”
Show real named people only back-facing, silhouetted, or through objects.
No essay renders a real person’s likeness.
TradeoffIt limits which compositions I can use.
Set reference strength by framing.
About 85% for close frontal shots, 72 to 78% for profile or three-quarter, 65 to 72% for crowds.
TradeoffOne more rule to apply to every image.
The rules, in one table.
About 84% less, by my estimate.
Estimate for 8 articles, typical case, at 2026 market rates. Includes my art-direction time.
Estimate for the same 8 articles, AI-assisted. Includes my art-direction time.
That is roughly 84% less. Across the scenarios I modeled, the range is 81 to 85% less.
These are my estimates from 2026 market rates. They are modeled, not measured.
What shipped, and what held.
- Image sets for 8+ long-form articles, each built section by section as a three-image arc.
- The same system supplied imagery for a second brand, Mission to Commission.
- Real reference photos held likeness. AI-generated references made it drift.
- Estimated cost: about 84% less than commissioned photography (modeled, not measured).
- Turnaround per article, before vs after: [turnaround per article, before vs after]
- Reader or client feedback: [reader or client feedback]
What I’d do differently.
I would build the real-reference library before writing the first prompt. The finding that real references held likeness came late, and it should shape the first step.
I would also write the review gate as a checklist on day one. It would have caught drift earlier.
I would capture turnaround time per article from the start, so the estimate could be paired with a measured number.
Assumptions and sources.
- Cost figures are my estimates from 2026 market rates, for 8 articles in a typical case.
- Both sides of the comparison include my art-direction time.
- Scenario range: 81 to 85% less.
- Sources for market rates: [add sources]