Standardize sci images

Standardize Scientific Images

Prepare comparable scientific images as a documented batch. Preserve raw files, apply only declared global operations, and fail when an honest scale bar or comparable display cannot be produced.

Start with the manifest

Read references/image_contract.md, then create a CSV manifest. At minimum it needs file, output_name, and um_per_pixel when a scale bar is requested. Add batch when images were acquired under different settings. The same comparison batch must share calibration and display settings.

python scripts/standardize_images.py manifest.csv \
  --outdir standardized --scale-bar-um 20

The command center-crops every image to a common pixel canvas and writes three practical derivatives: an unannotated display raster, a labelled preview, and an SVG panel whose faithful raster content remains embedded while scale bars and text stay editable vector/live layers. It also builds a montage and writes CSV/JSON audits containing source hashes and every operation. It never overwrites the source.

By default, the SVG physical width is set to the largest size that preserves 300 dpi. Use --panel-width-mm only for a known final slot; the command refuses a width that would fall below 300 dpi instead of inventing pixels.

Scientific integrity rules

• Treat raw acquisition files as immutable. Write derivatives to a new directory and record source SHA-256 hashes.
• Never invent a scale bar. Require calibration from image metadata, acquisition software, or an authoritative record.
• Do not infer calibration from another image unless the acquisition record proves they share it.
• Apply the same crop policy, intensity window, gamma, and LUT to images in the same comparison batch. Do not tune each experimental group separately.
• Default to no tone adjustment. If display adjustment is needed, declare fixed display_min, display_max, gamma, and lut settings in the manifest.
• Keep higher-bit-depth single-channel data at their native depth until display mapping. Require an explicit native-unit display window; never silently cast 16-bit data to 8-bit.
• Preserve a quantitative image/table separately. Display normalization is not a substitute for quantitative preprocessing.
• Do not erase, clone, selectively blur, locally enhance, or move biological structures.
• Do not resample by default. Equalize size by a common crop. If resampling is scientifically justified, perform it in a documented downstream workflow and update calibration.

Read references/integrity_and_layout.md before processing fluorescence composites, electron micrographs, unequal calibrations, or images used for quantification.

Layout rules

• Use one target canvas per comparison set, with identical output pixel dimensions and equal gutters.
• Use one physical SVG panel size per slot class and verify effective raster dpi at that final size.
• Use the same scale-bar length and placement within a comparable batch when it fits the shared field of view.
• Center the scale-bar label over the bar; do not right-align it to the bar endpoint. Keep this alignment and type size identical across the batch.
• Keep scale bars inside a safety margin and choose black or white from the local background for contrast.
• Do not add panel letters, serial numbers, per-image titles, or conclusions by default. Add only explicitly requested sample labels that are necessary to interpret the image.
• Keep captions, acquisition settings, and interpretation outside the image montage.
• Do not stretch images to make them fit. Crop consistently and record the crop box.
• Keep scale bars and their labels editable in the SVG delivery layer. Treat the underlying scientific image as faithful raster content, not as a fully vector-editable object.
• Judge scale-bar and sample-label size in the final panel or montage, not while zoomed into the source. Use one readable label hierarchy and baseline within each slot class.

Finish with the polish skill

Use $polish-sci-figures for final multi-panel grid assembly, font selection, whitespace control, overlap checks, and placement in the manuscript or slide. A visually aligned montage does not pass if calibration, processing provenance, or comparability is unresolved.

Read references/professional_basis.md for the image-integrity, publication-layer, calibration, metadata, and reporting sources behind these safeguards. The implementation is original to this repository.

Acceptance gate

Do not deliver if any raw file was changed, calibration is absent or guessed, comparison-batch display settings differ, crops remove different scientific regions without justification, outputs differ in size, a scale bar is clipped, incorrect, or too small at final size, or unrequested labels/titles/serial numbers appear.