Mirrors/photoprism
1
0
Fork 0
mirror of https://github.com/photoprism/photoprism.git synced 2026-07-17 16:49:04 +00:00
Code Issues Projects Releases Wiki Activity
photoprism/internal/ai/nsfw
History
Exact
Cathie Integra 3f096a592b Links: Use canonical trailing-slash form for website URLs
2026-06-24 17:20:30 +02:00
..
testdata Backend: Rename /internal/tensorflow to /internal/ai 2024-07-02 08:08:14 +02:00
.gitignore Backend: Rename /internal/tensorflow to /internal/ai 2024-07-02 08:08:14 +02:00
model.go Media: Harden Vision URL fetching against SSRF and DoS 2026-03-03 17:54:03 +01:00
nsfw.go Links: Use canonical trailing-slash form for website URLs 2026-06-24 17:20:30 +02:00
nsfw_test.go CI: Apply Go more linter recommendations to "ai/nsfw" package #5330 2025-11-22 11:33:28 +01:00
README.md Docs: Document NSFW flag gating and call-graph in package READMEs 2026-05-21 16:25:06 +02:00

README.md

PhotoPrism — NSFW Package

Last Updated: May 21, 2026

Overview

internal/ai/nsfw runs the built-in TensorFlow NSFW classifier to score images for drawing, hentai, neutral, porn, and sexy content. It is the default backend that powers the Type: nsfw model entry in internal/ai/vision and is the only NSFW engine that ships with PhotoPrism out of the box; operators can override it through vision.yml with an Ollama or OpenAI endpoint when they prefer to run NSFW detection on a remote LLM.

Where It Gets Called

The package itself only exposes the model loader and a thin scoring API (Result.IsSafe, Result.IsNsfw(threshold)). Two upstream callers wire it into the runtime:

  1. Upload handler — internal/api/users_upload.go. When PHOTOPRISM_UPLOAD_NSFW=false (default true), every accepted upload is screened by vision.DetectNSFW before indexing. Files that score above the threshold are deleted on the spot — they never reach originals/. When UPLOAD_NSFW=true, the upload path skips the check entirely.

  2. Index + vision-worker pipelines — internal/photoprism/index_mediafile.go, internal/workers/vision.go, internal/workers/meta.go. When PHOTOPRISM_DETECT_NSFW=true (default false), the indexer marks new photos as PhotoPrivate = true if the NSFW model flags them. Both code paths short-circuit when DetectNSFW() is false — the model is then neither loaded nor invoked.

Both flags are independent: you can reject uploads without flagging existing imports, flag existing imports without policing uploads, or both. The user-facing matrix lives at docs.photoprism.app/user-guide/ai/nsfw/.

Detection Through the Labels Model

When Type: labels is served by an Ollama or OpenAI engine and both PHOTOPRISM_DETECT_NSFW=true and PHOTOPRISM_EXPERIMENTAL=true are set, internal/config/config.go flips the package-level global vision.DetectNSFWLabels to true. The Ollama and OpenAI engine builders then swap their default label prompts for LabelPromptNSFW and the JSON schema generators add nsfw + nsfw_confidence fields, so NSFW classification piggybacks on the label-generation call instead of running as a separate inference pass.

When the shortcut is active, the labels-path check in index_mediafile.go (labels.IsNSFW(threshold)) can promote a photo to private without this package being touched. The dedicated TensorFlow model in internal/ai/nsfw is still used as a fallback whenever the labels path either does not run or does not return NSFW signals, and whenever vision run --models nsfw is invoked directly.

How It Works

  • Model Loading — Loads the NSFW SavedModel from assets/models/ and resolves input/output ops (inferred if missing).
  • Input Preparation — JPEG thumbnails (default size Fit720, see MediaFile.DetectNSFW) are decoded and transformed to the configured input resolution.
  • Inference & Output — Produces five class probabilities mapped into a Result struct for downstream thresholds and UI badges.

Threshold

vision.yml carries a Thresholds.NSFW value (default 75, range 0-100) that controls how confident the model must be before a picture is flagged. Lower values are more aggressive; higher values more permissive. The threshold applies to both the dedicated NSFW model and the NSFW fields returned via the label-generation shortcut.

Thresholds:
  NSFW: 75

Memory & Performance

TensorFlow tensors allocate C memory and are freed by Go GC finalizers. To keep RSS bounded during long runs, PhotoPrism periodically triggers garbage collection to return freed tensor memory to the OS. Tune with:

  • PHOTOPRISM_TF_GC_EVERY (default 200, 0 disables).
    Lower values reduce peak RSS but increase GC overhead and can slow indexing.

Troubleshooting Tips

  • Model fails to load: Verify saved_model.pb and variables/ exist under the model path.
  • Unexpected scores: Confirm the input resolution matches the model and that logits are handled correctly.
  • High memory usage: Adjust PHOTOPRISM_TF_GC_EVERY or reduce concurrent indexing load.
  • NSFW detection appears to stop working after switching labels to an LLM: Confirm both PHOTOPRISM_DETECT_NSFW=true and PHOTOPRISM_EXPERIMENTAL=true are set. Without both, the labels-path shortcut is disabled and only an explicit vision run --models nsfw (or another caller that goes through this package directly) will produce NSFW flags.

Related Docs