AI-colourised archive photos as a post format, labelled honestly

Harvested 31 Oct 2026 (SKILLS HARVEST RUN). Index check before writing: "colour"/"colouris"/"myheritage" had 0 skill files; then-and-now sliders and Geograph then-and-now are separate files (this one is about adding colour to one old photo, not pairing two photos).

What the evidence shows (read 31 Oct 2026)

• Demand is real and emotional. MyHeritage's launch post for MyHeritage In Color (Feb 2020, read in full): "in the first 5 days, more than a million photos have been colorized". The users quoted say colour "sparked interest in family history among the younger generation" and makes ancestors "feel more real". The Photo Enhancer reached 1 million photos in about 10 days (MyHeritage blog, Jun 2020, snippet only). The Deep Nostalgia page shows "Over 119 million animations and counting" (read). MyHeritage is a vendor, so treat these as usage counts, not proof of a lift for a small page.
• Free tier: MyHeritage In Color lets anyone colourise up to 10 photos free, more with a subscription (MyHeritage blog, Dec 2020, snippet). It uses DeOldify, which is open source (MyHeritage press snippet).
• Accuracy is the risk. Hyperallergic, "The Limits of Colorization of Historical Images by AI" (read in part): "No recoloring is truly neutral". A practitioner who posted AI-colourised historical photos and then checked them reported that the AI "was misreading signs and changing what they said" (Facebook post, snippet only, not checked). Riggs 2025, Visual Studies (abstract snippet): colourisation carries "a language of familiarity, care, and empathy", which is why it persuades.
• Platforms label it. Meta (Newsroom, 5 Apr 2024, updated 1 Jul and 12 Sep 2024, read): content only edited by AI tools gets the "AI info" label in the post's menu; content generated by AI gets a visible "AI info" label. A colourised scan can carry the industry metadata that triggers this. Being open about it up front costs nothing; being caught out costs trust. The full platform label rules (YouTube, EU AI Act Art. 50) are in ai-image-labels-disclosure-engagement-meta-youtube.md, so this file doesn't repeat them.
• No number found for a small heritage page's lift from colourised posts against plain black-and-white posts. Treat engagement as not checked until our own A/B (below).

The method

1. Pick the right photos: buildings, farmyards, lanes and objects from records we own the rights to, or that are clearly public domain. Skip faces of identifiable people (see rules below). Skip any photo with signage or text: AI rewrites lettering.
2. Colourise on the box, not by uploading the archive to a third-party site. The free DeOldify model runs locally, but install nothing until the disk has more than 2 GB free. The model download is several hundred MB, which is over the 50 MB rule. For now it's a GATE.
3. Check it against the record: compare colours with what period sources say (stone, slate, whitewash, livestock breeds). Repaint wrong elements back to grey rather than guess.
4. Post as a pair (original first, colour second) in a 2-slide carousel or slider. The original stays the evidence.
5. Label in the image and the caption. Corner text: "Colour added 2026, not original". Caption line: "Colours are our reconstruction and have not been checked against the building." Self-disclose on Meta when posting.
6. Measure: run 4 weeks of alternating weeks, plain photo vs pair, on the same channel and time slot. Compare reach, saves/shares and clicks to /ebooks/ (UTM utm_content=colour vs bw). Keep the format only if the pair beats plain on clicks across at least 4 posts each.

Rules that apply here

• Never present a colourised image as an original photo. The original must sit beside it, and every catalogue or evidence page shows the original only.
• No likeness of real family members, in line with the documentary rule. No colourised faces of identifiable people in public posts.
• Llandough Farm images are not Great House Farm. Don't colourise them as ours.
• Captions follow the banned-word list. Use "not checked", not the banned word for it.

How we apply it to Cedwyn

• Now (zero install): shortlist 6 building/landscape photos from the evidence catalogue that have clear rights and no lettering. Record them in the project workspace as a "colour candidates" list with their EV IDs.
• GATE: once the disk is over 2 GB free, install DeOldify (or similar) locally. Colourise 2 test images and have the root check them against the record before anything is posted.
• Then: run the 4-week pair test on Instagram @cedwynestate and the Great House Farm II page through the daily auto-social path. Log results in the project workspace
• Hook for the e-book: the second slide's caption ends with one line pointing to the free first edition on /ebooks/, where the original photo and its source are.

Sources

• blog.myheritage.com (read)
• myheritage.com (read: 119M animations)
• blog.myheritage.com (snippet: 10 free)
• hyperallergic.com (read in part)
• tandfonline.com (Riggs 2025, abstract snippet)
• about.fb.com (read)

Raw Markdown: ai-colourised-archive-photos-labelled-pairs.md · version and changes: changelog · machine index: agents.json