flowchart LR I["Images and metadata"] --> O["Operations and dependencies"] O --> P["Plan: phases, blocks and source regions"] P --> E["Execute ready tasks within resource limits"] E --> R["Images and diagnostic outputs"]
Processing large images with Blockflow
Blockflow is a Rust library for image processing when the data is larger than memory. Operations describe what they read and compute; a plan divides the work into blocks, and an executor runs the resulting dependencies with explicit storage and resource contracts.
Run a pipeline
Start with a small pipeline, then connect OME-Zarr images and choose blocks and memory limits.
Find an operation
Browse the operation catalog for filters, segmentation, registration, calibration and restoration. Entries distinguish planned execution from resident numerical helpers.
Understand the design
Follow an image through halos, planning, fusion and the scheduler.
From an image to executable work
A block owns an output core and may read a larger region around it. This extra region, the halo, makes local computations agree across block boundaries. The plan records those requirements before execution; scheduling changes when and where work runs while preserving the declared computation.
Follow one acquisition through the system
Three connected explanations use the same two-channel time series. Start with how acquisition dimensions become spatial volumes, follow one smoothing and thresholding run, then see which choices Blockflow makes automatically. The example keeps storage chunks, compute blocks and source halos distinct.
Practical boundaries
Blockflow is experimental and its API is evolving. Pin a Git revision for applications. Many operations support bounded image blocks, while some require whole axes, resident state or global solves. The operation and design pages explain those differences so that a small-image helper is not mistaken for an out-of-core algorithm.
See existing examples for complete applications and benchmarks for measured workloads and reproduction commands.