When we at Old Mentz Heritage Center take in photograghs. we scan them as 2400 dpi tiff files. The .tiff format is stable but dense and a typical cabinet card scan will be well over 100 mb in size. When I use in social media, I will run them through a image editor, resize and save as a jpeg file. So I asked ChatGPT what size and resolution is best for this type of identification project and it recommended 4500 pixels along the longest side at 300 or 400 dpi and saved as a jpeg. This will reduce the image from 100 mb to around 2 mb. This way it is dense enough for AI to read and small enough that you are not consuming data needlessly. If you have a large collection, manually converting every image from tiff to jpeg can take time, but maybe you focus on certain good examples?
I then switched to Gemini and put the general question to it.
“At our historical society, we have a collection of old family photos. A few are labeled but most are not. I am using one collection as a test example for teaching AI and its use in photo identification. I have scanned all the photos and uploaded them as 400 dpi jpeg files to our Google Photos account. Please treat me as a new comer and guide me through the process of extracting information and possible grouping of images. What can I expect to accomplish and how do I get there? What AI agents are best for this?”
1. What You Can Accomplish
By using AI, you aren’t just “sorting” photos; you are creating a searchable database. You can expect to:
-
Auto-Group Individuals: AI can find the same face across 50 years of photos, even as the person ages from a child to an adult.
-
Identify Context: AI can recognize objects (old cars, specific uniforms, types of architecture) to help you date an image.
-
Text Extraction (OCR): If any photos have writing on the back or signs in the background, AI can read and index that text.
2. Using Google Photos as Your First AI Agent
Google Photos is a “passive” AI agent—it works in the background without you needing to prompt it.
The “People & Pets” Tool
This is your most powerful asset for genealogical work.
-
Wait for Indexing: After uploading, Google’s servers take 24–48 hours to “look” at every pixel.
-
Access Collections: Go to the Collections tab and select People & pets.
-
Labeling: You will see groups of similar faces. Click a face and select “Add a name.” Once you name a person (e.g., “Great Aunt Martha”), Google will automatically suggest other photos it thinks are her.
-
The “Power of One”: If you have even one labeled photo of a person, the AI uses that as a “seed” to find them in the rest of the unlabeled pile.
Searchable Visual Content
You don’t need to tag everything. Try searching for:
-
“Wedding” or “Uniform” or “Bicycle” — the AI identifies these objects automatically.
-
“Text: Smith” — Google will search for any scanned handwriting or printed text containing that name.
AI Agent Best For… How to Use
Gemini (or ChatGPT-4o) Contextual Analysis Upload a photo and ask: “Based on the clothing and car model, what decade is this?”
Ancestry / MyHeritage Genealogical Matching These specialized AIs compare your faces against millions of other public family trees.
Daminion / Tonfotos Local Archiving If your society wants to keep data offline, these tools run facial recognition on your local computer.
The “Historical Detective” Prompt
Here is the Gemini suggested prompt that you can use as a model.
“I am [working with a historical society] or [attempting] to identify this [family] photograph. Please act as a visual historian and provide a detailed analysis to help us narrow down the date and context. Please look for and report on the following:
Fashion & Grooming: Analyze clothing styles, hat shapes, collar types, and hairstyles. What decade do these suggest?
Technology & Objects: Identify any visible technology (cars, bicycles, cameras, farm equipment) or household items.
Environmental Clues: Look at architectural styles, street signs, or vegetation that might suggest a specific geographic region or season.
Photo Characteristics: Based on the grain, tonal range, and physical condition (scratches, borders), what can you infer about the photographic process used (e.g., tintype, cabinet card, early film)?
Summary Estimate: Based on all clues, provide a likely date range (e.g., 1910–1915) and your reasoning.
Please be specific about any ‘anchor points’ you find—items you are certain belong to a specific era.”
Tips for Better Results:
-
Mention the Location: If you know the general area where the family lived (e.g., “This family was from rural Ohio”), add that to the prompt. It helps the AI distinguish between, for example, a New York City fashion trend and a Midwestern one.
-
The “Back” of the Photo: If there is any writing on the back of the photo, scan that too! You can upload both images and ask the AI to “Synthesize the visual clues from the front with the handwriting style and text on the back.”
-
Zoom In: If there is a tiny detail (like a license plate or a newspaper headline), crop that area and upload it as a second image for the AI to “read.
I then uploaded this photo and asked for clues. 
Gemini came back with this overview;
This is a fantastic example for your historical society test case. By combining the visual clues from the front and the studio information on the back, we can pinpoint a very narrow window for this photograph.
1. Visual Historian Analysis
-
Fashion & Grooming: * The Mother: She is wearing a bodice with “leg-of-mutton” sleeves (puffed at the shoulder and tight at the wrist), which reached their peak size and popularity in the mid-1890s. Her high, stiff collar and tightly pulled-back hair with small curls are classic mid-to-late Victorian.
-
The Father: He sports a high-buttoning sack coat and a thick, well-manicured mustache, typical of the 1890s. His detachable wing collar and patterned cravat suggest a formal studio portrait.
-
The Children: The girls’ dresses feature gathered yokes and ruffled shoulders, consistent with mid-1890s children’s wear. The boy’s double-breasted tweed jacket is a staple of young men’s fashion from 1892 to 1898.
-
-
Technology & Objects: * The father has a pocket watch chain (fob) visible across his waistcoat.
-
The furniture is a “photographer’s prop” chair with plush upholstery and fringe, common in studios of this era.
-
-
Photo Characteristics: * This is a Cabinet Card, indicated by the heavy cardstock mount and the studio branding.
-
The sharp detail and silver-gray tones suggest a gelatin silver or collodion print, which were standard by the late 19th century.
-
2. Studio Detective Work (The Back of the Card)
The reverse of the card provides the most concrete geographic and chronological “anchor points”:
-
The Studio: A.L. Cramer Photographic Studios.
-
Locations: The card lists addresses in Scranton (309 & 311 Lackawanna Ave), Carbondale, and Jermyn, PA.
-
Historical Context: Records indicate that A.L. Cramer was active in the Lackawanna County area during the 1890s. Specifically, the “Lackawanna Ave” address in Scranton was a high-traffic professional district during the region’s coal and rail boom.
3. Summary Estimate
Likely Date Range: 1894–1897
Reasoning: The sleeve volume on the mother’s dress is the most reliable “time stamp.” Sleeves began to puff significantly around 1892, reached their maximum size in 1895/1896, and began to shrink again by 1898. The combination of the sleeve size and the specific studio branding makes 1895 a very strong candidate for the year this was taken.
I then uploaded the two images that I used for my ChatGPT conversation and asked it to compare them to the family protrait.
About 1882
About 1885
This is a perfect example of how AI can help “connect the dots” across different time periods in a collection. By analyzing the facial structures and the specific studio equipment, we can identify these individuals as members of the same family, roughly 10–15 years apart.
1. The Cross-Identification (Visual Grouping)
By comparing the features, we can reasonably identify the children from the A.L. Cramer group portrait in their younger years:
-
The Girl: This is very likely the older girl standing in the middle-back of the family portrait. Note the identical deep-set eyes, the shape of the mouth, and the distinct ear shape.
-
The Boy: This is likely the boy from the family portrait as a toddler. The facial proportions—specifically the spacing of the eyes and the slight “cleft” appearance of the chin—match the older boy standing behind the father in the family shot.
2. Dating the Early Photos (The “Before” Era)
While the family portrait dates to roughly 1895, these two individual portraits are significantly older.
The Girl
-
Fashion: She is wearing a dark “Garibaldi” style bodice with decorative buttons and a lace-trimmed Peter Pan collar.
-
Clue: Notice the striped stockings and high-button boots. These were extremely popular for young girls in the late 1870s to early 1880s.
-
The Studio Prop: She is leaning on the exact same fringed velvet chair seen in the boy’s photo. This confirms they were taken at the same studio, likely on the same day.
The Boy
-
Fashion: He is wearing a “Knickers” suit with a wide belt and a small bow tie. This style for young boys was prevalent in the early 1880s.
-
Condition: The double-border line around the edge of the photo is a common design for Carte de Visite (CDV) or early Cabinet Cards from the 1880–1885 period.
3. The “Anchor” Analysis
4. Teaching Moment: How the AI Groups These
If you were using Google Photos or a facial recognition agent (like Tonfotos), here is how the AI would process this:
-
Feature Mapping: The AI looks at the “Golden Triangle” (the distance between eyes, nose, and mouth). Even as a child grows, these ratios remain relatively stable.
-
Prop Detection: An advanced AI would recognize the unique fringe and woodwork on the studio chair in images 010 and 011, automatically grouping them as being from the same “session.”
-
Metadata Suggestion: Once you label the boy in the 1895 photo as “William Richfield,” the AI would likely flag the toddler in image 010 as a “Potential Match.”
But I had uploaded the wrong image of the girl, using the later one instead ot the one that was the twin of the boy. So I uploaded all three and asked Gemini to compare again.
The young girl, about 1882
This set provides a perfect “laboratory” for testing how AI identifies individuals across different ages and sessions. By looking at the physical evidence in the frames and the subjects themselves, we can build a very strong chronological sequence.
1. The “Session A” Evidence (c. 1882)
Your hunch is spot-on: the boy and girl were almost certainly taken on the same day.
-
The Matching Borders: Both images share an identical double-line border design on the card mount.
-
The “Same Moment” Props: The girl in 013 and the boy in 010 are using the exact same fringed chair and holding the same small, white knitted or beaded purse.
-
The Accessory: Note the girl’s dress in 013 has a distinct plaid/checkered band at the bottom.
2. The “Session B” Evidence (c. 1885)
The older girl represents a slightly later moment in time.
-
The Border Change: The card mount has changed to a single, thicker rounded-corner border, which became more common as the 1880s progressed.
-
Subject Growth: This is almost certainly the same girl, but she has aged roughly 2–3 years. Her face has lengthened, and her hair is styled differently, though she is still wearing the same high-button boots.
-
The Callback: She is holding the same white purse and leaning on the same fringed chair. This suggests the family returned to the same photographer (or the same studio location) a few years later.
Teaching Insight: “Prop Consistency”
In AI training, this is called Contextual Association. Even if the AI was unsure about the girl’s face, it would find a 100% match on the fringed chair and the beaded purse. In archival work, these “constant” objects are often the best way to prove that a box of loose photos all came from the same local studio.
So now I have a family group matched with photos of the kids at a younger date. Do I know who they are? Nope. But at least I have grouped some photos removing them from this part of the process. So far, both ChatGPT and Gemini have proved to be very capable of conducting this deep dive research.