PhotoScribe v1.7.0

Captions for different destinations, ready-to-paste social posts, and much better advice on which model to run.

Added

• Captions for different destinations. A new Outputs card chooses what the caption written into the photo is for. Catalogue is the existing factual style and stays the default. Stock writes literal titles, 25 to 45 keywords ordered most important first, no brand names or subjective praise, and uses the editorial caption format when a photo shows identifiable people or brands. Flickr writes a gallery-style description and tags, since Flickr reads them from the file on upload.
• Ready-to-paste posts for Instagram, Facebook, Threads / Bluesky and Mastodon, plus alt text. Tick the ones you want, then press Write posts on a photo in Results to write them for that photo, without touching its title, caption or keywords. Most photos in a batch never get posted, so this keeps Generate fast; tick Write posts for every photo during Generate if you want them for the whole batch. Each post has a copy button and a character count that turns red past that platform's limit. Posts are never written into the photo: those platforms strip embedded metadata on upload, and hashtags don't belong in a catalogue.
• An option to write posts in the first person, as the photographer sharing their own photo, limited to reactions to the photo itself. Read posts before you use them, as you would anything a model writes.
• Copy buttons on Title, Caption and Keywords in Results.
• A warning when generation is crawling. Below 8 tokens a second, the log explains that the model is almost certainly too big for the memory available.

Changed

• Recommend Model now suggests Gemma 4, sized to leave real headroom. It previously named Gemma 3 and counted 70% of a Mac's memory as usable, ignoring Lightroom and everything else running. 12GB graphics cards now get the 12B rather than the 4B, and Macs are advised to use an MLX build in LM Studio.

Model choice matters more than any setting

A model that can read images isn't automatically a good fit. One that only just fits in memory gets read back from disk as it runs. Measured on a 32GB M2 Max with the same photos: Gemma 4 12B ran at 33 tokens a second, about 8 seconds a photo, while a 27B and a 26B-A4B crawled at 3 to 6, and didn't write noticeably better captions. If PhotoScribe is slow, it's almost certainly the model. Use Recommend Model. The README has the full table.