MediaPipe and NumPy: distinguish dependency constraints from binary import failures

Scope and symptoms

Installation can fail during resolution, or a completed installation can fail while loading an extension. Read the actual MediaPipe distribution rather than applying a historical constraint to every release.

Search fragments; wording and context vary:

mediapipe
_ARRAY_API not found

Source-supported context

The MediaPipe Python setup guide documents supported desktop platforms and python -m pip install mediapipe; it does not prescribe one NumPy version for all releases. A 2024 issue reported numpy<2 for MediaPipe 0.10.15. The current PyPI 1.0.1 metadata and wheel list declares numpy without that upper bound and offers platform-specific wheels, so the old issue cannot be used as a current universal pin. NumPy's import-error guide identifies _ARRAY_API not found and numpy.dtype size changed as possible downstream binary incompatibilities; the first non-NumPy frame in the traceback helps identify the failing package.

Distinguish these cases

1. The exact installed MediaPipe release and another package declare incompatible NumPy ranges, or pip cannot find a wheel matching the interpreter/platform.

2. Installation completes but a downstream native extension fails at import with _ARRAY_API, dtype size changed or another ABI error; the failing extension may be a different package in the same node stack.

3. Imports work and the remaining failure concerns a node resource, image representation or model operation.

Suggested diagnostic sequence

This sequence is editorial guidance, not a diagnosis of your machine.

1. With the Python executable that starts ComfyUI, record MediaPipe, NumPy and node versions, interpreter/architecture, pip's selected wheel and either the resolver error or full traceback. Separate installation from runtime import.

2. Read Requires-Dist for the exact installed release and run a dependency check. Compare the ranges of all affected consumers; do not paste numpy<2 into a shared environment merely because it appears in a 0.10.15 discussion.

3. If no supported range intersects, isolate the workloads. For an ABI error, follow NumPy's traceback advice to identify the first non-NumPy package, then test its maintained compatible release or documented rebuild in a recoverable copy. Avoid unidentified binaries and broad reinstall loops.

4. After the exact failing import works, run the node author's minimal non-private input and regress other retained NumPy consumers. A later model-path or image-format failure is a distinct stage, not evidence to repeat downgrades.

Completion checks

Dependency metadata is consistent, relevant native imports work and both the minimal node and retained workflows pass.

Limits and cautions

A broad numpy declaration in current metadata does not prove every native extension in a shared ComfyUI environment works with every NumPy build. Platform wheel availability also varies; a passing metadata check does not establish ABI or functional compatibility.

Source review: 2026-09-26. No package installation, native import, MediaPipe task, GPU inference or user-environment repair was executed.

Original sources

• MediaPipe Python setup · 2026-09-26
• MediaPipe issue 5676: historical NumPy constraint · 2026-09-26
• MediaPipe PyPI release files · 2026-09-26
• NumPy import-error troubleshooting · 2026-09-26

Related guides

• ModuleNotFoundError in ComfyUI even though the package is installed
• Keep a reproducible ComfyUI repair record without exposing secrets