mito_aff_setup53-2d_16nm-ft3

A mito_affinities model for electron microscopy: it predicts mito_aff_z, mito_aff_y, mito_aff_x (raw logits (apply sigmoid / threshold at 0)), from 16 nm voxels to 16 nm voxels, in blocks of 1×556×556 voxels in and 1×468×468 out (fp16).

Finetuned in the browser with cellmap-flow 0.3.0 from cellmap/onnx-web-models/setup_53_2d (Unet3dIso (2D-collapsed)): its whole network was trained on annotations painted in the browser for 20 epoch(s), to a final loss of 0.1 (2026-10-26). The annotations were painted on https://janelia-cosem-datasets.s3.amazonaws.com/jrc_mus-liver/jrc_mus-liver.zarr/recon-1/em/fibsem-uint8.

Open it

• Try it in the browser: cellmap-flow opens with the model chosen; connect any volume to run it.
• Open it on the dataset it was trained on.
• Anywhere cellmap-flow takes a model (Add a model in the dashboard's Models tab, in the browser or on a server): ackermand/mito_aff_setup53-2d_16nm-ft3/mito_aff_setup53-2d_16nm-ft3.

mito_aff_setup53-2d_16nm-ft3/manifest.json says the model's voxel sizes, blocks, channels and input normalization in cellmap-flow's format, and mito_aff_setup53-2d_16nm-ft3/mito_aff_setup53-2d_16nm-ft3.onnx is the ONNX graph.