TomoJAX support matrix

Available workflows and their entrypoints. “Supported” identifies an implemented and tested path, not a guarantee of accuracy for every scan. Alignment remains experimental. See installation and measurement scope.

No-flag tomojax align and tomojax recon use the Python automatic policies. Their recorded outcome is conditional: reported limitations are not accuracy certificates. Choose an explicit mode/method when supplying expert solver settings.

ScopeCopy Anchor Link

NXtomo and generic TIFF/HDF5 frame loading requires recorded or explicitly supplied detector pixel sizes and view angles. External NXtomo/Data Exchange use their format-defined axial parallel model. If the pitch is recorded but the rotation centre is absent, loading estimates detector-u from a few corrected rows with opposite-view gradient correlation and records inferred provenance; vertical zero defines the free axial origin. Insufficient coverage/contrast requires explicit detector=. Laminography requires a recorded tilt/axis or geometry=. Nikon and Zeiss files lacking dark calibration require darks=; zero is only an explicit choice for already corrected counts. SkyScan additionally requires flats= (frames or one level) rather than deriving it from bit depth. Compressed TIFF crops decode only their intersecting strips/tiles. In folders, TIFF-suffixed entries are image storage, not acquisition text sidecars. TIFF line-integral import requires explicit --geometry, --pixel-size and detector_center = [u, v] in --config; laminography also needs tilt_deg. For raw frames without an explicit grid, detector binning also reduces output resolution. Axial-parallel crops follow the measured row window; other models retain their output box. File/caller-supplied grids and Scan.binned() retain their chosen sampling.

Missing values produce an error naming the override, not centred unit pixels or a 30° tilt.

Workflows outside the table above are research or expert diagnostics.

Acquisition metadataCopy Anchor Link

load_frames recognises these raw exports by content, even with a renamed file. A partial detector record is retained for inspection, not completed with a zero rotation centre. Missing axial centres may be estimated from data; an unavailable pitch or failed estimate requires detector=/geometry=. Physical pitches below are normalised to millimetres.

These readers do not turn an estimated motor position into a calibrated rotation centre, or infer pixel pitch from image dimensions. Data Exchange exports lacking facility labels remain Data Exchange, not a guessed beamline. This support does not establish support for every native TOMCAT/SYRMEP layout. Real-file coverage includes ESRF NXtomo converted by nxtomomill, TOMCAT/SYRMEP Data Exchange, Diamond I12 HDF5 and K11 GDA, Nikon TIFF acquisitions, Zeiss TXRM, SkyScan logs/TIFFs, Diamond I13 compressed TIFF folders with explicit calibration, and TomoJAX NPZ/NXtomo round trips. Native EDF/Bliss inputs must include every file named by the selected sequence or parameter file: missing inputs are errors. These checks do not establish support for native TOMCAT GigaFRoST. Missing calibration does not disable content recognition: supply flats= (frames or one level) and darks=. NXtomo/Data Exchange angular fields use their format-defined degree convention when units are absent; generic motor datasets without units require explicit angles=.

HDF5 (view, u) sinograms retain lazy frame access and are presented with a singleton row as (view, 1, u), including during NXtomo inspection/validation. PyHST multi-frame, cone, helical, binning and distortion modes are rejected; pass a .par explicitly rather than relying on directory discovery to choose it.

SkyScan logs supply distances, camera pitch/binning, optional resize and the vertical optical-axis line, but not a calibrated horizontal centre. Supply detector=/geometry= plus explicit flat/white and dark calibration.

Zeiss reads ImageInfo/Angles, object-plane pixelsize, XrayGeometry, cone distances, both alignment/*-shifts and AutoRecon/CenterShift; missing calibration requires geometry= rather than zero image shifts. The embedded reference uses ReferenceData/DataType; separate XRM flat/dark files use their own ImageInfo/DataType. Both are read lazily. MultiReferenceData/ImageN references use their recorded exposure durations and timestamps; the same count-rate correction handles separate XRM references.

Nikon detector pitch/offsets and source distances come from the .xtekct fields; its axis requires either the recorded modern offset/roll/tilt or both legacy COR endpoints. Vendor reconstruction dimensions, pitches and ROI offsets remain source metadata, not an applied reconstruction region: TomoJAX uses its field-of-view grid unless the caller supplies grid=. Both TIFF axes and the rotation/calibration signs are normalized together. Default angles use InitialAngle and AngularStep, not .ang or _ctdata.txt; an explicit angles= schedule must be in TomoJAX's object-angle frame.

Numerical and accelerator scopeCopy Anchor Link

Volumes use (x, y, z) array order and projections use (view, v, u). Voxel and detector spacings must use consistent physical length units. Forward projection integrates over physical ray length. FBP returns attenuation in the volume's units; its backprojection normalisation differs from the Euclidean transpose used by iterative solvers. See accuracy and performance evidence for the measured cases, backend choices, and limitations.

For choosing an alignment mode and checking its result, see the alignment guide.