Command-line reference
Run imcluster --help to see the reference for the installed version.
Positional arguments
INPUTS...Optional image files, directories, or text manifests to process. Inputs may be omitted when
--cachenames an existing imcluster cache.
Main options
--cache PATHPreserve processing results in a reusable Parquet cache.
--gallery PATHPreserve the standalone HTML cluster gallery at this path.
--no-openDo not launch the generated gallery in the default browser.
--dino-version [auto|2|3]Select the DINO generation. Automatic mode prefers an accessible or cached DINOv3 model and otherwise falls back to DINOv2.
--arch [vit|convnext]and--size [tiny|small|base|large|huge|max]Select a model preset. DINOv2 ignores
--archand supportssmall,base,large, andmax(Giant). DINOv3 uses both options.--model TEXTUse an arbitrary compatible Hugging Face model instead of the preset.
--device [auto|cpu|cuda|mps]and--batch-size INTEGERControl model inference hardware and batching.
--recursiveand--max-images INTEGERControl input discovery.
--clustering [spectral|dbscan|hdbscan|kmeans|agglomerative|hierarchical]Select clustering behavior. Fixed-count methods use
--n-clusters; DBSCAN uses--dbscan-eps. DBSCAN and HDBSCAN share--min-samples. The default iskmeans.--reduce [none|umap|tsne|pca]Optionally reduce DINO feature dimensions before clustering. Reduced vectors are cached. The default is
umap; use--reduce noneto disable it.--reduction-dims INTEGERSet the requested output dimensionality for PCA and UMAP. The default is
50. t-SNE is capped at three dimensions.--nameGenerate and cache descriptive cluster names from representative cached thumbnails using a multimodal LLM.
--llm TEXTand--llm-temperature FLOATSelect the llmloader-compatible naming model and its sampling temperature. Defaults are
gpt-5.6-lunaand0.2.--llm-api-key TEXTPass a provider API key for cluster naming. Prefer the provider’s environment variable to avoid exposing a secret in shell history or process listings.
--in-group-size INTEGERand--out-group-size INTEGERSet the maximum in-cluster and contrasting outside thumbnails sent to the naming LLM. The defaults are
10and0respectively.--thumbnail-width INTEGERand--thumbnail-height INTEGERSet the maximum embedded thumbnail dimensions.
--evaluate PATH(alias:--expected PATH)Evaluate clustering against a CSV with
filename,classcolumns and print Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), and optimally matched clustering accuracy (ACC).--metric PATHWrite NMI, ARI, and ACC scores to a CSV. Requires
--evaluate.
Cache options
--force replaces every cached stage. --force-features,
--force-cluster, and --force-thumbnails recompute
individual stages and their required downstream results.