One of the strange quirks of fate is that I have inherited thousands of old photos. The majority are unidentified. I guess my relatives decided they couldn't identify the people either, and I was a willing recipient of the whole pile. About 4,500 of the original photos came from my great-grandmother, Margaret Godfrey Jarvis Overson. She was a professional photographer, as was her father. Her collection went to the University of Arizona Special Collections Library in Tucson, Arizona (see
The old photos are all in the public domain.
I am also a professional photographer, though not as active as I have been in the past. See my
Now what about using AI to repair or enhance some of these old photos? You can see that the photo above has several problems. The photo is torn and has a significant crack or line. Parts of the photo are obscured, and the boy's faces are washed out with little detail. The first issue is AI itself. AI does not repair or enhance the photos. It takes the original photo and regenerates or generates a similar copy, only with the additional information that it's missing from the original photograph. Here is a simple example of Google Gemini's Nano Banana's enhancement effort for this photo with a simple prompt to repair and enhance.
There are a few differences between this "enhanced" photo including parts of the photo that were missing have been restored, and the photo has been cropped slightly, apparently across the bottom. Most people would agree that it is an improvement over the original photo.
However, for historical purposes, it is not the original photo. What happens if I use a very developed, specific prompt giving instructions to Gemini about the actions it should take with regards to this old photo?
Here is the prompt that I am going to use:
Act as a master photographic conservator and senior digital restoration specialist specializing in archival preservation and Adobe Photoshop. Your objective is to analyze a historical photograph and generate an actionable, surgical digital restoration plan that prioritizes ethical conservation over cosmetic recreation.
### Core Guiding Principle
Adhere strictly to the Principle of Minimal Intervention: repair only what is compromised to stabilize the image, leaving 100% of authentic original detail, paper texture, and natural film grain intact. Never recommend global blurs, aggressive noise reduction, automated AI over-smoothing, or destructive layer merges.
---
### Step-by-Step Assessment & Repair Workflow
1. Medium & Baseline Analysis
- Identify or estimate the photographic process (e.g., Daguerreotype, Tintype, Albumen, Gelatin Silver, Cyanotype) and approximate era.
- Catalog critical intact elements (e.g., facial feature micro-details, authentic paper grain, midtone density) that must be strictly protected from alteration.
2. Damage Inventory & Categorization
Map the damage across three distinct categories:
- Structural Defects: Cracks, tears, missing emulsion/substrate, fold creases, punctures.
- Chemical & Surface Deterioration: Silver mirroring, foxing/mold, water stains, fading, chemical discoloration.
- Optical / Scanning Artifacts: Dust, scratch lines, scanner glare, sensor noise.
3. Non-Destructive Photoshop Layer Architecture
Provide an explicit, bottom-to-top Photoshop workflow organized into dedicated layer groups:
- Group 1: Base & Registration (Locked original capture layer, working duplicate, reference guides).
- Group 2: Structural Repairs (Sample All Layers healing/cloning, custom texture patching, brush modes).
- Group 3: Texture & Surface Cleanup (Frequency Separation: High/Low parameters, localized Clone Stamp / Healing Brush settings, blend modes).
- Group 4: Density & Contrast Correction (Targeted Curves adjustment layers, localized Black & White/Channel Mixer adjustments, Luminosity masks).
4. Surgical Tool Specification Table
Present a structured table mapping each damaged area to its precise repair method:
| Location / Feature | Damage Type | Photoshop Tool / Technique | Layer Settings & Blending Mode | Preservation Constraint |
|---|---|---|---|---|
| [e.g., Subject's cheek] | [Scratch] | [Spot Healing (Proximity Match)] | [Dedicated Blank Layer, Sample All Layers] | [Avoid adjacent pore texture loss] |
5. Quality Control & Verification Protocol
- Specify a "Difference Mode" or toggle-check inspection routine to verify that zero alterations have spilled into undamaged zones.
The advantage of using the prompt is that more of the original photo is preserved, and there are no apparent additions other than the original photo's missing areas
But I suspect that those people who are using AI to enhance photos would be expecting something completely different from of this photo "restoration." My long, specific prompt would not do the changes. It explained in great detail the reason why the changes could not be made for historical reasons. if I go back to ignoring the specialized prompt, and simply give some additional details to the photo, outlining the changes needed to the first copy of the photo you might get something like this:
My prompt consisted of the phrase, "Take out the large crack, take out the dark spots all over the photo, and enhance the background." I think that most people would agree that this was a better enhancement of the photo, but any connection to the historical, original photo has been almost completely severed. the background has been completely changed. Details have been added, especially to the floor and the faces. my guess is that as long as this technology is available, it will be used by most people, including most genealogists, to enhance their photos, especially very damaged photos like the one I use here as an example.
The most important issue is ensuring that the original photograph is preserved—both as a physical artifact and as an unedited digital representation. Whenever an enhanced version is shared online, it should be clearly marked as AI-enhanced, with the original copy provided alongside it.
As a quick note, none of the photos I put up on FamilySearch have been enhanced or changed in any way by AI.
No comments:
Post a Comment