Every built-in workflow node

Every node a workflow can use, in the editor's categories. Inputs and outputs are port names; a port's type is its name (image, detections, classifications, depth). A parameter with no value shown has no default: it is either optional or, for a file, chosen when you build the workflow.

Model nodes take a variant, one of the family's variants on its model page. Nodes that take text take a list of phrases, one per line in the editor, each used exactly as written.

The same list is available from a running server at GET /workflow/nodes, and in Python from mozo.workflow.catalogue(). Both are generated from the nodes themselves, so they always match what runs.

Input

Node Inputs → outputs Parameters What it does
read_media none → image source, stride=1, start=0, count, rate Read an image or a video: one frame for the first, every frame for the second.
read_annotations image → detections path The detections save_annotations wrote for this image, read back.

Output

Node Inputs → outputs Parameters What it does
save_image image → none path=output, format=.png Write each image to its own file.
save_video image → none path=output.mp4, fps Write every frame that reaches this node to one video file.
save_annotations image, detections → none path=annotations.jsonl Write what was detected, one line of JSON per item, as each item finishes.

Detect

Node Inputs → outputs Parameters What it does
yolov8 image → detections variant=nano, threshold=0.5 YOLOv8 by Ultralytics: real-time object detection.
yolov11 image → detections variant=nano, threshold=0.5 YOLO11 by Ultralytics: real-time object detection.
yolov12 image → detections variant=nano, threshold=0.5 YOLO12 by Ultralytics: attention-centric real-time object detection.
yolov26 image → detections variant=nano, threshold=0.5 YOLO26 by Ultralytics: NMS-free detection, and instance masks on the seg variants.
rfdetr image → detections variant=nano, threshold=0.5 RF-DETR by Roboflow: NMS-free transformer detection, under a permissive licence.
rfdetr_checkpoint image → detections checkpoint, architecture=nano, labels, threshold=0.5 RF-DETR from a checkpoint of your own: a model you trained, run like a published one.
owlv2 image → detections text (a list of phrases), variant=base-ensemble, threshold=0.1 OWLv2 by Google: find anything you can name, with no training.
grounding_dino image → detections text (a list of phrases), variant=tiny, threshold=0.3 Grounding DINO by IDEA: open-vocabulary detection from a phrase.

Segment

Node Inputs → outputs Parameters What it does
sam3 image → detections text (a list of phrases), variant=sam3, threshold=0.5 SAM 3 by Meta: masks for every instance of a concept you name.
sam2_boxes image, detections → detections variant=tiny SAM 2 by Meta: a mask for each box it is given, the rows otherwise unchanged.
edgetam_boxes image, detections → detections variant=edgetam EdgeTAM: SAM 2 made small: a mask for each box it is given, the rows otherwise unchanged.
sam3_boxes image, detections → detections variant=sam3 SAM 3 by Meta: a mask for each box it is given, the rows otherwise unchanged.

Edit

Node Inputs → outputs Parameters What it does
moebius image, detections → image variant=general, seed=0, dilate=0 Moebius: remove what you point it at and repaint the gap. Wire a segmenter into detections.
ben2 image → image variant=base, stretch=True, refine=False BEN2: cut the background away, keeping a soft edge. Takes no prompt and no boxes.

Pose

Node Inputs → outputs Parameters What it does
vitpose image, detections → detections variant=small ViTPose++: the joints of everyone you point it at. Wire a detector into detections.

Classify

Node Inputs → outputs Parameters What it does
clip image → classifications text (a list of phrases), variant=base CLIP by OpenAI: score an image against phrases you make up.
siglip2 image → classifications text (a list of phrases), variant=base-224 SigLIP 2 by Google: zero-shot classification against phrases you make up.

Read

Node Inputs → outputs Parameters What it does
easyocr image → detections variant=english EasyOCR by JaidedAI: find text in an image and read it.

Depth

Node Inputs → outputs Parameters What it does
depth_anything_v2 image → depth variant=small Depth Anything V2: how far away everything is, from one photograph.

Track

Node Inputs → outputs Parameters What it does
track detections → detections tracker=bytetrack, lost_after=1.0 Give each object the same tracker_id from frame to frame, so it's counted once however many frames it is in.

Transform

Node Inputs → outputs Parameters What it does
rotate image → image angle=90.0 Rotate the image about its centre, in degrees.
flip_horizontal image → image none Mirror the image left to right.
flip_vertical image → image none Mirror the image top to bottom.
crop image → image left=0, top=0, right, bottom Cut a rectangle out of the image. Right and bottom default to the image's own edges.
resize image → image height=720 Scale the image to a height, keeping its shape.
to_grayscale image → image none Drain the colour, keeping three channels so the result is still an image.
rotate_with_detections image, detections → image, detections angle=90.0 Rotate the image and carry the detections round with it.
flip_horizontal_with_detections image, detections → image, detections none Mirror the image left to right, and the detections with it.
flip_vertical_with_detections image, detections → image, detections none Mirror the image top to bottom, and the detections with it.
crop_with_detections image, detections → image, detections left=0, top=0, right, bottom Cut a rectangle out, keeping the detections that fall inside it.
align_by_keypoints image, detections → image, detections first=left_eye, second=right_eye, target_angle=0.0, detection_index=0 Rotate so that two named keypoints sit at a chosen angle: how a face is straightened.
crop_around_detections image, detections → image padding=0.0 Cut out one image per detection.
pad_detections detections → detections padding=0.1, reference=shorter Grow every box by a fraction of its own size.
bbox_from_keypoints detections → detections keypoints Redraw each box to enclose its keypoints. Unset uses all of them.

Adjust

Node Inputs → outputs Parameters What it does
enhance_clahe image → image clip_limit=2.0, tile_size=8 Lift local contrast, which brings detail out of shadow without blowing out the rest.
auto_contrast image → image cutoff=1.0 Stretch the histogram to the full range, ignoring the brightest and darkest few percent.
gamma_correction image → image gamma=1.0 Brighten or darken without clipping. Below 1 brightens, above 1 darkens.

Annotate

Node Inputs → outputs Parameters What it does
draw_boxes image, detections → image thickness, color Draw a box around each detection.
draw_labels image, detections → image position=top_left, font_scale, padding=6, text_color=#FFFFFF, background Write each detection's class and score beside it.
draw_masks image, detections → image opacity=0.5, color Shade each detection's mask over the image.
draw_polygons image, detections → image thickness, color Outline each detection's mask rather than filling it.
draw_keypoints image, detections → image radius, thickness, color, show_names=False, min_confidence=0.0 Mark each detection's keypoints.
draw_skeleton image, detections → image thickness, color, min_confidence=0.0 Join each detection's keypoints into a skeleton.
blur_regions image, detections → image kernel_size, padding_percent=0.05 Blur whatever was detected: faces, plates, anything you would rather not publish.
pixelate_regions image, detections → image pixel_size, padding_percent=0.05 Pixelate whatever was detected, where a blur would look like a mistake.

Prompts read back the same from every model

The prompted nodes are sam3, owlv2, grounding_dino, clip and siglip2. Each result's class_id is the index of the phrase that found it (for CLIP and SigLIP 2, of the phrase it scored), and its class_name is that phrase. So results read the same whichever model produced them. A list with an empty entry is refused rather than shortened, because shortening it would change every later phrase's class_id.

Write phrases as short noun phrases: "hard hat", not "a photo of a hard hat". The models that locate things find nothing for a caption. On the example photograph, SAM 3 and Grounding DINO each found six or more people for "person" and none for "a photo of a person", while CLIP scored both the same.

Boxes into masks

sam2_boxes, edgetam_boxes and sam3_boxes add a mask to every box they are given, in one call per image, and keep every row as it came: same order, class, name, score and box. A row whose mask comes back empty keeps its box with an empty mask. An image with no boxes never runs the model. sam3_boxes after a sam3 node on the same image reuses that node's image encode.

Three graphs cover most annotation jobs:

Job Graph
Find objects by name, with masks read_media → sam3
Find boxes, then outline them read_media → grounding_dino → edgetam_boxes
Give boxes you already have masks read_media + read_annotations → edgetam_boxes

rfdetr_checkpoint runs an RF-DETR you trained: choose the checkpoint file, the architecture it was trained from, and its class names one per line in class-id order, or leave the names empty to use the ones the checkpoint carries.