Nodes

move is the one node that changes where an asset’s bytes live — a trash folder, a storage pool, a library, or an S3 bucket for cold storage. It refuses to overwrite, works out whether a destination is on another disk before it starts copying, and only removes the original after the destination has been verified. assign is its counterpart for curation: it files an asset into a collection or a library and never touches the file at all.

md5, sha256, sha512, chunk-hash run in parallel over any file. Hashes are the basis for exact deduplication and for skipping already-processed assets.

fingerprint computes a perceptual multi-sector fingerprint using the video4j-fingerprint library. It survives transcoding, resolution changes and minor edits, which makes near-duplicate detection possible at scale via vector similarity search.

facedetect finds faces in images and video with the InspireFace library, returning bounding boxes, facial embeddings and the optimal focal point for cropped thumbnails. Its detections port carries one element per face, so facedescription, scene-layout and dominant-color all work per face simply by connecting to it.

objectdetect finds objects — people, vehicles, animals, furniture — in images and video frames with a YOLO model, reporting each as a labelled box with a confidence score. The classes it can report come from the model’s own vocabulary, so choosing the model is choosing what the library can be searched for. Its labels port lists the distinct classes found and connects straight to tag, which turns every asset into something searchable by what is in it.

sam2 answers the question detection cannot: not where a thing is, but which pixels are it. Given the boxes objectdetect already found, it returns each one as an exact outline; given nothing at all, it finds every distinct region in the picture by itself; given a video, it follows an outline from frame to frame as the subject moves. That boundary is what separates a product from its backdrop, crops to a person rather than to a rectangle around one, and turns "how much of this shot is sky" from an impression into a number.

thumbnail renders contact-sheet previews from video frames and source images and uploads them to Loom as asset binaries.

metadata reads what a file already says about itself — the photographer’s title, caption and keywords, the camera and exposure, the licence and credit line, and where the shutter was pressed — out of EXIF, IPTC, XMP, companion .xmp files, PDF and Office properties and audio tags, and normalises all of it onto one Dublin Core shape. It records which original field supplied each value, so nothing has to be trusted blindly, and it needs no model and no GPU. Its text output carries the authored prose alone, ready for translate, sentiment or llm.

tika uses Apache Tika to pull the full text out of any supported document format. It is the companion to metadata: one gives you the words, the other the properties.

scene-detection splits video into scenes with an optical-flow detector, producing cut timestamps that can drive per-scene thumbnails or clips.

ocr extracts text from images and document pages with Tesseract; vlm transcribes document pages with a vision-language model such as olmOCR, which keeps the layout — tables, columns, formulas — that plain glyph recognition flattens; whisper transcribes audio and video soundtracks with the Whisper model.

captioning generates natural-language descriptions of both images and videos with a vision-language model — images from a still frame (SmolVLM), videos from their temporal content (Qwen2.5-VL), optionally as a per-scene timeline — and llm enriches metadata through an LLM provider such as llama.cpp or vLLM.

imagegen generates a new image with a diffusion model behind a small HTTP sidecar — either from a text prompt or by remixing the asset’s own image — and writes the PNG to the worker’s local cache.

videogen generates a short video clip with a diffusion model behind a small HTTP sidecar — either from a text prompt or by animating the asset’s own image as the opening frame — and writes the MP4, with a synchronised audio track, to the worker’s local cache.

watermark stamps a logo or mark you supply onto an asset at a relative position, so one setting works across a library of mixed resolutions. Stills are composited in-process; video goes through the ffmpeg overlay filter, which re-encodes the picture but copies the audio through untouched. The original file is never modified — the marked copy is a new artifact.

image-manipulation straightens, crops and reframes a picture in one pass — applying its EXIF orientation so it is upright, cutting it to a fixed region or one built around the faces connected to its detections port, forcing a target aspect ratio (padding vertical pictures with a blurred enlargement of themselves rather than black bars), and bounding the result’s size. No model or external tool is needed, and the original file is never modified.

tts turns text connected to its text port — an llm summary, a caption, or a translated transcript — into narration audio, using Orpheus/Kartoffel for German and Kokoro for English behind a small HTTP sidecar.

sentiment scores the tone of text the pipeline already produced — connect document text, OCR output, a caption or an LLM answer to its text port — as positive, neutral or negative with a signed polarity value, using dedicated German and English models behind a small HTTP sidecar.

translate turns text the pipeline already produced into another language — connect a transcript, document text or an OCR result to its text port and the translation comes out on translation, ready for tts to speak. One node translates into one language, so several languages means several nodes; each translation is stored under its own language and stays separately retrievable. Documents longer than the model’s context window are split at paragraph and sentence boundaries, never mid-sentence, and rejoined.

guard checks content against a harm taxonomy — connect a transcript, document text, an OCR result or a caption to its text port, or point it at image assets — and answers safe or unsafe with a confidence score and the categories that fired. Its safe output is a branch, so a publishing or upload step connected behind it runs only for the items that came back clean, while flagged assets are still fully processed and their verdicts stored for review. Three families of open guardrail model are supported behind the one node — Meta’s Llama Guard, Google’s ShieldGemma and IBM’s Granite Guardian — and all of them are normalised onto one shared vocabulary of harm categories and one comparable score, so changing the model does not mean rewiring the pipeline. Llama Guard 4 and ShieldGemma 2 read pictures directly; the rest screen text.

depthmap works out how far each part of a photo is from the camera, from the photo alone, and scene-layout uses that to say how the detected things relate — which is in front, which is behind, which one hides part of another. Together they turn a flat list of boxes into a scene you can caption accurately and search spatially.

dominant-color works out which colours a picture is actually made of — the whole frame, a fixed region, or every box arriving on its detections port — and reports each one as HEX, RGB, HSL and CIELAB alongside a name a person would use, in English and German. It needs no model and no GPU, so it can run over an entire archive cheaply.

consistency validates that stored metadata — hashes, dimensions, codec information — still matches the file on disk, flagging inconsistencies for review or remediation.

gdrive-source and onedrive-source read media straight out of Google Drive, OneDrive and SharePoint, without anyone having to copy it to a disk or a bucket first. Both are differential: a re-run asks the drive what changed and processes only that, so watching a large drive costs a single request when nothing has happened. And because a file in a drive keeps its identity when someone renames it or files it away, both report a move as a move rather than as a deletion followed by a fresh upload — so the metadata you have already built up stays attached to it.

tag attaches tags from rules you configure over what the pipeline already worked out — a quality metric, a transcript, a list of colour names. It matters because of where a tag lands: most analysis results sit on the asset where you can read them but not search them, while a tag becomes part of the asset’s search document as soon as it is written. A vocabulary can be pinned to an allowed list so automatic tagging cannot litter the shared tag list, and the node only ever removes a tag it can prove it wrote itself.

filter sorts items into buckets you configure and grows one output port per bucket, plus an other port for everything else. The port is the branch: anything wired to a port that carried nothing is skipped for that item, and the skip carries on down that branch — so one node routes a pipeline any number of ways without a condition repeated at every step. Today it sorts by language, decided by a language model.