NSFW Detection

PhotoPrism can automatically flag pictures as private when an image-classification model considers them unsafe for work. It can also reject such files during web upload so they never enter the library. NSFW detection is opt-in and is disabled by default. It is intended mainly to help administrators of shared instances keep adult content out of their libraries without having to review every upload manually.

Configuration Options

Two independent config options govern the runtime behavior. Both are off by default:

Config Option	Effect
PHOTOPRISM_DETECT_NSFW	When true, photos detected as NSFW during indexing or a photoprism vision run are marked as private. When false (default), NSFW signals are ignored even if the underlying model returns them.
PHOTOPRISM_UPLOAD_NSFW	When false, the web upload dialog rejects files that the NSFW model flags as unsafe (the rejected file is deleted before indexing). When true (default), uploads are accepted regardless and any NSFW flagging happens later during indexing per DETECT_NSFW.

The two options are independent: you can reject uploads without flagging existing imports, flag existing imports without policing uploads, or both.

Which Model Detects NSFW?

The model used for NSFW detection follows the same vision.yml mechanism as labels, captions, and faces. If no Type: nsfw entry is configured, PhotoPrism uses the built-in TensorFlow NSFW model. You can override it just like any other model — for example, to point at an Ollama or OpenAI endpoint trained to score NSFW content:

Models:
  - Type: nsfw
    Model: <your-vision-model>
    Engine: ollama
    Service:
      Uri: ${OLLAMA_BASE_URL}/api/generate

Once the dedicated NSFW model entry is configured, it runs whenever PHOTOPRISM_DETECT_NSFW is true and the indexer or photoprism vision run --models nsfw invokes it.

The built-in TensorFlow NSFW model is small and fast and ships with every PhotoPrism image. There is no need to configure an LLM-based replacement unless you want to.

Using a Labels Model

When Type: labels is served by Ollama or the OpenAI API, PhotoPrism can ask the model to include NSFW classification in the same response, avoiding a second inference pass. This shortcut is experimental and gated by two environment variables that must both be set to true:

• PHOTOPRISM_DETECT_NSFW
• PHOTOPRISM_EXPERIMENTAL

If either is false, the label-generation prompt will not ask for NSFW fields, so the LLM response cannot trigger NSFW flagging — even if an image contains NSFW content. The following matrix summarizes the behavior when Ollama or OpenAI is configured for labels and the user runs photoprism vision run --models labels:

Detect NSFW	Experimental	Labels Prompt	Outcome
false	false	Default labels prompt (no NSFW fields)	NSFW detection does not run.
false	true	Default labels prompt (no NSFW fields)	NSFW detection does not run.
true	false	Default labels prompt (no NSFW fields)	NSFW detection does not run via labels; the dedicated NSFW model only runs when nsfw is explicitly included in the run's models (e.g. --models labels,nsfw).
true	true	NSFW-aware prompt (nsfw, nsfw_confidence)	LLM returns NSFW fields; photos above the configured threshold are flagged as private. The dedicated NSFW model still runs as a fallback when nsfw is included in the run's models.

If you switched from TensorFlow to an Ollama or OpenAI labels model and noticed that NSFW detection stopped working, the most likely cause is that one of the two flags above is missing. Enable both PHOTOPRISM_DETECT_NSFW=true and PHOTOPRISM_EXPERIMENTAL=true, or include nsfw in the explicit --models list so the dedicated NSFW model runs alongside labels.

NSFW Threshold

The Thresholds.NSFW value in vision.yml controls how confident the model must be before a picture is flagged. The threshold accepts an integer between 0 and 100 and defaults to 75. Lower values are more aggressive (more pictures flagged); higher values are more permissive. The threshold applies to both the dedicated NSFW model and the NSFW fields returned through the label-generation prompt.

Thresholds:
  NSFW: 75

Running NSFW Detection Manually

To re-evaluate the existing library after enabling or tuning NSFW detection, run the vision worker with the nsfw model in a terminal:

docker compose exec photoprism photoprism vision run --models nsfw

Add --force to override existing private flags, and restrict the run to a subset of the library by appending a search filter (same syntax as the --vision-filter config option), for example:

docker compose exec photoprism photoprism vision run --models nsfw --force public:true

Learn more ›