Curator Settings
AI-tag model
Used for artwork tagging, precis generation, and bio drafting.
Runs this same model on both cards so a queued batch drains roughly twice as fast. Needs the model's second-GPU registry entry to exist and be loaded; otherwise runs single-GPU and says so in the log.
Which prompt profile bulk AI-tag runs use. Add a new version to this instance's prompt file, select it here, then re-run items filtered on the old aitagrun/ tag to compare — every run stays in the item's history.
How many queued AI-tag jobs run together on each card. Each card runs its own backend with its own KV cache, so the practical ceiling is that cache divided by the size of one request — and with both GPUs on, the total in flight is double this number. Image-only libraries (generated, judge) send a near-constant ~3k-token prompt and comfortably sustain 8. Art defaults to 2 because its prompt also carries the evidence bundle and artist context, several times the size and far more variable item to item. Set it to 1 for one job at a time.
This instance's own load requirements, sent with every AI-tag call. If the model is resident with different values it is reloaded to these — only once idle, never mid-request. Leave blank to accept whatever is loaded (the registry supplies defaults on a fresh load); enter 0 to clear a value. Only set these if this instance genuinely needs a different pool than the model's default — matching declarations never cause a reload.
Reads the prompt and completion sizes this instance's own runs actually recorded, and sizes the pool for concurrency x (prompt + max_tokens) — what a sequence is allowed to use, not what it happened to use. Leave the box blank to size for the jobs-at-once setting above. Nothing is saved until you press Save Preferences.
Image Similarity
Sets the starting closeness and result limit for the high-level Similar action. You can adjust closeness live on the results page. Fast search uses cached local visual features without the GPU; the optional SigLIP deep-search button adds cached semantic image embeddings.
Gallery Display
The ★ badge in the corner of each tile. Ratings themselves are unaffected — set them from the grid (1–5, 0 clears) or the image view either way.
AI-tag prompt shelf
Every system prompt that has actually run in this library, grouped automatically — no saving needed. Apply makes one active; Name labels it; View shows the full text. Model is a separate choice (filter by aitagmodel/* tags).
Biography model
Artist and movement biographies are drafted from source text — no image is sent — so this list includes text-only models the AI-tag selector hides. Leave it on "Same as AI-tag model" to keep the previous behaviour.
VLM Context Limits (characters)
These caps are applied at prompt construction time. They do not affect stored extraction text.
Unified (Tags + Précis)
Artist Bio
Movement Bio
Prompt Review
Read-only view of prompt templates used by the VLM stack.