The Memory Preservation Standard
What we do, what we don’t do, and why.
Last updated: 2026-04-27
AI photo restoration as a category has earned skepticism. Generic apps smooth faces, invent detail, swap identities. We built our pipeline against that critique. The document below is the record of how.
The Three Doctrines
Identity-Sacred Invariant
The person in the photograph stays the person. We do not invent faces. We do not let a model drift toward an averaged version of someone you know.
Every restoration is checked against this rule before delivery. When the check finds a problem, the photograph routes to a human Conservator instead of shipping with quality compromised.
Quality Promise
The first photograph and the hundredth photograph receive the same care. We do not run a faster pipeline for cheaper tiers. We do not skip steps for volume orders. The customer paying for one photograph and the customer paying for an archive both get the same restoration.
Conservation Voice
We use words like preserve, restore, honor, finish. We do not use words like magic, reimagine, perfect. The language reflects the work. Restoration is craft — not mystery, not invention, not a promise of an ideal that does not exist.
Conservator Review
Some photographs are hard. Heavy damage. Complex restoration. Ambiguous identity. These route to a human reviewer rather than ship with a quality compromise.
Today, that human is our founder. As we grow, it will be a small team trained to the same standard. The customer never sees the difference: the queue is invisible to the funnel, and the delivered photograph passes the same checks regardless of whether AI or a human did the final work.
No Invented Detail
A list of what we will not add to a restoration:
- faces or facial features that aren’t in the original
- people who aren’t in the original
- clothing patterns, jewelry, or accessories that aren’t there
- era-mismatched detail (1940s clothing rendered with 2020s textures)
- hands or fingers when restoration would require speculation
- hallucinated text or logos
- color information beyond what era and context support
The Nine Failure Modes — and how we prevent each
The professional restoration field has named nine specific ways consumer AI restoration goes wrong. We address each directly.
1. Halo and ghosting
Halos and ghost edges appear when an upscaler treats faces like generic textures. Our pipeline runs face-aware restoration that recognizes facial regions and treats them as their own preservation problem.
2. Texture mismatch
Old photographs have grain. Smoothing it away makes the result look like a modern phone capture pretending to be old. We preserve the photographic grain that anchors the photograph in its era.
3. Facial reconstruction drift
Generic models drift toward average faces — your grandmother starts to look like someone else’s grandmother. Identity-preservation guardrails catch the drift; outputs that show measurable identity shift route to human review before shipping.
4. Plastic skin
Plastic, airbrushed skin is the signature of consumer AI restoration. We preserve identifying detail — pores, lines, shadows — because those are what make a face the face you remember.
5. Temporal inconsistency
A 1940s portrait should not return looking like 2020s digital art. Our pipeline anchors restoration to the era of the original through colorization decisions, contrast handling, and stylistic choices that respect the source.
6. Depth and contrast loss
Many AI restorers flatten the tonal architecture of the original — the dark areas brighten uniformly, the highlights blow out. We preserve the depth and contrast curve of the source photograph.
7. Synthetic noise patterns
Some restorers add fake film grain to make outputs look “authentic.” We do not insert noise. Whatever grain the original photograph carries is what the restoration carries.
8. Hallucinated details
The damage-check stage compares the restoration to the original on a checklist of known invention modes. If anything appears in the restored photograph that wasn’t in the original, the photograph is rejected and routes to human review.
9. Hand and finger failures
Hands are where AI restorers most often invent. Extra fingers, missing fingers, hands that don’t connect to wrists. When a photograph needs hand reconstruction, we route it to a human Conservator rather than risk the failure mode.
What We Do Not Do
Some product directions are permanent rejections. We will not build these even if asked.
- Photo animation. We do not bring still photographs into motion through generative video. The line from preservation to invention crosses there.
- Destructive single-output sharpening with no undo. The customer must always be able to recover the original and the prior restoration version.
- Auto-clustering of faces across a customer archive without consent. Even if technically simple, the privacy boundary is firm.
- Generative face reconstruction beyond the original photograph. The Identity-Sacred Invariant means we do not build a face from a name. We restore the face that was already in the frame.
Print Quality — What We Promise, What We Don’t
Our pipeline delivers digital files suitable for digital sharing and casual prints up to 11×14 inches. The output is faithful at that size and below.
For larger fine-art prints or museum-grade output, the work happens human-to-human. We scan the original photograph at high resolution, hand-finish the restoration, and prepare it for archival printing. That is consultation work, not website work. Reach out to discuss a print consultation.
We do not list a single price for that service because every consultation is its own scope. We list the digital pipeline price honestly because every digital order runs the same code path.
Damage-Check — How It Works
Every restoration is evaluated against a checklist of known failure modes before delivery. The check looks at identity preservation, hallucinated detail, texture authenticity, era consistency, and hand reconstruction.
When the check flags a problem, the photograph is held for Conservator review instead of shipping. The customer receives the higher-quality result after that review — not an automated output that may have failed silently.
Most consumer AI photo apps do not have a step like this. We chose to build it because the alternative is a photograph that looks fine until you look closely, and then never looks right again.