Importing bioimage.io Models

bioimage.io is a community repository of pretrained deep-learning models for bioimage analysis, each published with a Resource Description File (rdf.yaml) describing its architecture, inputs/outputs, and (optionally) pre/post-processing requirements.

Instead of manually filling in a Stardist, UNet, Cellpose, or YOLOv5 command by hand, you can point EVAnalyzer at a downloaded model’s rdf.yaml and have it pre-fill an AI Segmentation step for you.

Importing a Model

Download a model package from bioimage.io and extract it to a local folder. The folder must contain an rdf.yaml (or rdf.yml) file.
In the steps panel, click Add step - or + between two steps where the segmentation step should go.
In the command picker, click Import bioimage.io model….
Select the model’s rdf.yaml file in the file picker.

EVAnalyzer parses the file, detects the model architecture, and inserts a pre-configured pipeline step at that position - no separate download step is needed for the RDF metadata itself.

Supported Model Types

Model	Detected when…
YOLOv5	The RDF’s name, description, or tags mention “yolo” - checked first, since a YOLO model card may also mention or cite Cellpose
Cellpose	The RDF’s name, description, or tags mention “cellpose”
Stardist	The RDF’s name/description/tags mention “stardist”, or it has a `config.stardist` block
UNet	Fallback for any other segmentation model

Whatever is detected from the model card is filled in automatically:

Cellpose - model path, and Input channels read from the RDF’s input tensor shape (defaults to 2 if it can’t be determined).
Stardist - model path, plus Probability threshold / NMS threshold if present in the RDF’s config.stardist.thresholds block.
YOLOv5 - model path. Every model class i is written as segmentation class i + 1; set Class mapping to choose. If the model was trained on downscaled images (YOLOv5 shrinks them to 640 px by default), set Image scale to 640 / training image size, e.g. 0.3125 for 2048 px images - the import notes remind you of both.
UNet - model path, and the output layout: a 2-channel mask+boundary output is mapped to Independent Channels (foreground = channel 0, boundary = channel 1); a multi-class softmax output assumes the foreground is the last channel.

After the step is inserted, a dialog lists any caveats about what was assumed - review these and adjust the step’s parameters if they don’t match your model:

Channel count or output layout could not be determined from the RDF, so a default was used.
The model declares input preprocessing (e.g. normalization) - EVAnalyzer does not apply this automatically; add equivalent steps (e.g. Intensity Transform) before the AI Segmentation step.
A boundary-aware UNet model was detected - add a Connected Components step afterwards to separate touching objects.
The model’s weights are hosted remotely - download the weights file yourself and point Model path at the local file.

Requirements

The RDF must describe a model resource (not a dataset or notebook) and ship TorchScript weights (weights.torchscript, or the deprecated weights.pytorch_script key). A model that only provides raw pytorch_state_dict weights cannot be imported directly - convert it to TorchScript first.
Both RDF spec versions 0.4 and 0.5 are supported, including RDF 0.4 output shapes given relative to an input tensor (reference_tensor + scale/offset), which some boundary U-Nets such as joyful-pig use.
A model that only lists ONNX weights can’t be run - the import tells you so. Add a TorchScript export to the RDF (for YOLOv5: export.py --include torchscript).
Only the four segmentation architectures above are recognised. Classification or other model types will not be configured correctly.
Importing a model does not require AI/CUDA support to be installed - it only writes pipeline configuration. Running the resulting step still requires a build with the ai feature enabled (see AI Models).