LoginSign Up
← Back to all sparks
P

PhotoPrism

COLLAB
Velocity3.8

AI-powered photo app for browsing, organizing and sharing photo libraries

www.photoprism.app ↗

PhotoPrism swapped in an ONNX vision pipeline, then opened the library to 360° media at 16K.

photo-libraryself-hostedonnx-visionmedia-formats360-photoslocal-llm
◆Current state
Two substantial releases in ten weeks moved the product on both axes it cares about. May replaced the legacy Pigo face detector entirely with an ONNX-based pipeline, made vision.yml accept model identifiers from Hugging Face, Ollama and OpenAI-compatible catalogs, and modernised media handling with Vulkan transcoding, a native HEIC/AVIF reader and layered TIFF and PSD support. July added interactive equirectangular 360° photos and video, raised the thumbnail and video ceiling to 16K, and put an inline multi-page PDF viewer in the app.
◆Where it's heading
The AI work has moved from bundled TensorFlow models to a pluggable pipeline where the model catalog is configuration, not a build artifact — Ollama, OpenAI and Hugging Face identifiers all resolve through the same file. In parallel the media layer keeps absorbing formats other libraries push out to external tools: PSD, layered TIFF, PDF, and now 360°. PhotoPrism is positioning as the thing you point at a mixed archive, not just a photo folder.
◆Prediction
Expect the format expansion to continue and the vision configuration to keep broadening its provider surface, with the accessibility section added in July likely to grow into a fuller pass.

◆Recent moves

  1. 5d ago

    360° media, 16K thumbnails and an inline PDF viewer

    ⚡ SPARK

    The format-expansion thread reaching its widest point: 360° media becomes a first-class type, resolution ceilings rise to match, and documents are readable in place. It follows May's media-layer rework, which is what made handling this class of file practical.

    View source ↗
  2. 2mo ago

    Disk-space threshold halts indexing before volumes fill

    A service release whose most useful addition is operational: indexing, importing and uploading can stop automatically below a configurable free-space threshold. Also fixes VAAPI transcoding under FFmpeg 8 and hidden-stack recovery — cleanup after May's media-layer rework.

    View source ↗
  3. 2mo ago

    ONNX pipeline replaces Pigo; vision models become configuration

    ⚡ SPARK

    The release that decoupled PhotoPrism's vision stack from its build. Everything since — the format work in July, the model-catalog flexibility — depends on the pipeline and media-handling changes made here.

    View source ↗
  4. 4mo ago

    Ollama configuration via env vars, broad reliability fixes

    Adds OLLAMA_BASE_URL and OLLAMA_API_KEY plus better fallback handling for caption generation with reasoning models, alongside edge-case fixes in indexing, Places, folder albums, thumbnails, WebDAV and OIDC. Explicitly framed as stabilisation ahead of the May release.

    View source ↗
  5. 8mo ago

    Batch Edit dialog, Ollama and OpenAI caption generation

    Batch metadata editing arrives after being promised for two releases, and the Ollama integration extends from vision into generating captions and labels, with OpenAI added alongside and custom TensorFlow models supported. The groundwork for treating the model as a user choice rather than a bundled asset.

    View source ↗
  6. 1y ago

    Ollama models in Vision AI and an Adjust Location dialog

    The first appearance of Ollama in the vision service, plus a map dialog for correcting photo coordinates and album deletion from the toolbar. Batch editing is flagged as in final development, which it was — it shipped four months later.

    View source ↗