ezmi2d

ezmi2d is an experimental reader for the line-oriented MI drawing format used by HP ME10 and PTC Creo Elements/Direct Drafting. Its parser and reference resolver are written in Rust and exposed to Python through PyO3.

The public API provides typed geometry, annotations, and part structure while retaining the complete, byte-preserving raw scan:

import ezmi2d

drawing = ezmi2d.read("drawing.mi")
modelspace = drawing.modelspace()

print(drawing.version, drawing.units, drawing.extents)
print(drawing.encoding, drawing.encoding_source)

for entity in modelspace.query("LIN ARC CIR"):
    print(entity.id, entity.mi_type, entity.property_id)

line = drawing.entitydb[10]
if isinstance(line, ezmi2d.Line):
    print(line.start, line.end)

for text in modelspace.query("TEXT"):
    print(text.text, text.origin, text.height, text.font_name)

for spline in drawing.query("FIL BSPL"):
    print(spline.id, spline.mi_type)

for sheet in drawing.sheets:
    print(sheet.name, sheet.child_part_indices)

for occurrence in drawing.iter_instances():
    print(occurrence.path, occurrence.part.name, occurrence.world_transform)

for dimension in drawing.dimensions:
    print(dimension.measurement, dimension.formatted_text, dimension.text_position)

for hatch in drawing.hatches:
    print(hatch.pattern, hatch.boundary_loops)

Matplotlib preview

Install the optional plotting dependency and render a decoded fixture to PNG:

pip install "ezmi2d[plot]"
python examples/plot_mi.py tests/data/geometry.mi -o geometry.png

The plotting helper can also be embedded directly:

import matplotlib.pyplot as plt
import ezmi2d

drawing = ezmi2d.read("drawing.mi")
fig, ax = plt.subplots()
ezmi2d.draw(drawing, ax=ax, expand_instances=True)
ax.legend()
fig.savefig("drawing.png", dpi=160)

ezmi2d.draw(Document) renders every decoded part definition once by default; pass expand_instances=True to traverse nested/shared instances and apply their transforms, or pass one part to inspect its definition. This is a semantic diagnostic preview, not a style-faithful MI renderer. Arc.ccw is True for the corpus-verified orientation code 0; unverified codes are None and are skipped with a warning rather than guessed.

read() accepts a path or bytes-like object, including the verified gzip-compressed MI envelope. Point and property pointers are resolved after all records have been decoded, so forward references work and dangling or wrong-type references produce source-located diagnostics rather than aborting the whole drawing.

Text is decoded strictly. The reader checks an explicit encoding= override, a UTF-8 BOM, MI 3.20-or-newer version metadata, and a #~1 ENCODING: declaration before conservatively inspecting known text fields in older files. Supported canonical encodings are utf-8, shift_jis (including cp932 and windows-31j aliases), and hp-roman8:

drawing = ezmi2d.read("legacy-japanese.mi", encoding="cp932")
text = drawing.query("TEX")[0]
print(text.text, text.text_bytes, text.content_value.encoding)

Invalid or undecidable text is never decoded with replacement characters. TextValue.text remains None, TextValue.raw_bytes retains the source, and a source-located diagnostic identifies the bad byte.

The lower-level lossless scanner remains available when inspecting unsupported records or developing new entity decoders:

scan = ezmi2d.scan("drawing.mi")
for record in scan.records:
    print(record.section_number, record.record_type, record.raw_bytes)

packed = ezmi2d.scan("drawing.bi")
print(packed.format.compression, packed.container_size, packed.source_size)
assert packed.container_bytes != packed.source_bytes

For compressed inputs, every source span addresses the decompressed logical MI stream exposed as source_bytes. The exact caller input remains available as container_bytes. Streaming decompression is bounded independently by ScanLimits.max_file_size, max_decompressed_size, and max_compression_ratio.

The CLI emits a raw structural summary or JSON suitable for corpus inspection:

ezmi2d inspect drawing.mi
ezmi2d inspect drawing.mi --json --records

Current scope

• legacy MI 2.10 global metadata, drawing extents, units, and serialized transform values
• #~6 part ownership and an ezdxf-style modelspace(), query(), and entitydb
• typed P, LIN, ARC, FIL, BSPL, CIR, and TEX entities; point references are resolved
• De Boor evaluation for BSPL, including rational weights when a verified layout records them
• typed dimension (DANG, DCHMF, DDIA, DRAD, DSGL) references, measurement, formatted text, placement, DDA/DTF style, and linked DTV tolerances
• drawable LED vertices, SYML components, and associative HAT patterns/boundaries through typed COC, PFA, and HAPP records
• typed Affine2D, nested/shared ASSE occurrence traversal, composed child-to-parent transforms, multi-sheet instances, root parts, and DOCU_SHEET links
• strict UTF-8, Shift_JIS/CP932, and HP Roman-8 text decoding with explicit override support
• typed PSTAT, ASSP, DTA, DTF, DDA, DLA, DAF, and HAPP property models, while retaining unverified numeric enumerations in source order
• stable diagnostics for duplicate IDs, bad records, dangling pointers, wrong pointer types, table-of-contents mismatch, and Phase 1 structural problems
• UnsupportedEntity fallback with its original RawRecord; no addressable record is silently discarded merely because its semantic decoder is unavailable
• bounded line, section, and |~ scanning with read-only access to every logical MI byte range
• streaming gzip decompression with container-size, expanded-size, ratio, truncation, checksum, trailing-data, and concatenated-member guards

Graphic source semantics are exposed as color, linetype, lineweight, resolved property tables, and LAYER: attributes. Legacy display_values remains a compatibility tuple whose fourth value is the property count, not visibility. The modern extra header value is retained as visibility_value; visibility stays None until that code is independently verified. ARC/FIL retain raw orientation and separately expose ccw: bool | None; an unknown code is never converted to False. Modern variable-prefix B-splines expose display_values=None and retain that prefix as prefix_values instead of guessing a style layout. closed, periodic, rational, and weights are presence-aware; the currently verified layouts do not establish those meanings and therefore return None, distinct from an explicitly recorded False. For TEX, the serialized 3x3 transform, translation, rotation, width factor, mirror state, alignment, primary/alternate fonts, line spacing, and multiline content are named. Every post-ID field remains available through values. Dimension and annotation layouts are named only where the product corpus has a self-consistent reference or coordinate role; remaining fields, B-spline prefixes, and assembly relationship values stay lossless bytes. Writers are not implemented. gzip-wrapped product-generated compressed MI is supported by content signature. zlib-wrapper, ZIP, UNIX compress, and UNIX pack signatures are recognized and rejected as unsupported instead of being decoded by assumption. The available genuine sample is a compressed 2D MI member inside a Creo bundle, not a separately exported standalone .bi; compatibility with every Drafting/ME10 .bi generation is therefore not claimed.

See the Python API reference, MI format research, and sample corpus guide. Third-party drawings are not distributed in Git, sdists, or wheels.

The opt-in corpus suite currently exercises 19 MI 2.10 drawings containing 10,166 points, 4,030 lines, 1,059 arcs, 353 fillets, 1,196 circles, 6 B-splines, and 57 text entities. The paired DXF tests compare line endpoints, arc center/start/end points, circle centers/radii, drawing extents, and text content/insertion/height to seven decimal places. Fillets exactly match the additional DXF arcs. Stored B-spline interpolation points and every vertex of the corresponding DXF polylines are evaluated against the decoded curves.

It also verifies a product-generated MI 3.40 / UTF-8 compressed member from a public Creo bundle: 87,506 compressed bytes expand to 393,805 logical bytes. Its 853 lines, 187 arcs, 88 B-splines, 76 circles, 57 texts, 88 typed annotations, 11 hatch contours, 9 hatch associations, 25-part hierarchy, and sheet association produce identical semantic models and diagnostics from compressed and expanded input. All 46 dimensions resolve their source geometry and point references; all 16 symbols resolve three graphic components. One COC member is a retained unsupported PLN record.

Distribution

Release automation builds cp310-abi3 wheels for Linux x86-64 and ARM64, macOS universal2, and Windows x64. Artifact-only smoke tests cover Python 3.10 and 3.14, and the sdist is rebuilt and installed in a clean job. See docs/releasing.md for the complete release gates and Trusted Publishing setup.

Development

uv sync --locked --extra dev
uv run maturin develop
uv run pytest
cargo test --workspace
uv run python benchmarks/benchmark_pipeline.py tests/data --warmup 0 --repeat 1 \
  --json-output /tmp/ezmi2d-benchmark.json --markdown-output /tmp/ezmi2d-benchmark.md

Parser fuzzing uses nightly Rust and cargo-fuzz:

cargo +nightly fuzz run raw_scan
cargo +nightly fuzz run semantic_read

To fetch and run the separately licensed corpus comparison:

./scripts/fetch_external_samples.sh
uv sync --locked --extra dev --extra corpus
uv run --extra corpus pytest tests/python/test_external_corpus.py