compare-stages

name	compare-stages
description	Diff two USD stages by prim/mesh/vertex/material count, file size, and validator summary. Use for before/after optimization comparisons.
version	1.0.0
allowed-tools	Shell, Read
metadata	{"author":"NVIDIA Corporation","tags":["usd","diff","comparison","validation"]}

Compare Stages

Produce a structured diff between two USD stages — typically the original asset and the optimized output. Reports what changed at multiple levels: file size, prim/mesh/vertex/material counts, and (optionally) validator findings.

What this skill covers

• Usage — arguments and modes.
• Step 1 — collect metrics from both stages.
• Step 2 — present the comparison table.
• Step 3 — (optional) diff validator summaries.
• Step 4 — (optional) prim-level diff.

Companion skills: run-operations (produces the optimized stage), run-validators / interpret-validators (validates before and after), inspect-asset (single-stage inspection).

---

Usage

Argument	Meaning
<before.usd>	Path to the original stage.
<after.usd>	Path to the optimized / modified stage.
--validators	Also diff validator findings (requires saved artifacts for both).
--prims	Show prim-level adds/removes (can be verbose on large stages).

Workflow:

1. Pass <before.usd> and <after.usd> as positional arguments.
2. Run the metric-collection script (Step 1) under a Python that can import pxr.
3. Use the captured JSON to render the comparison table (Step 2).
4. Optionally pass --validators to diff validator issue counts between the two stages (Step 3).
5. Optionally pass --prims to execute the prim-diff helper (Step 4).

If either path is missing, ask:

"Please provide the paths to the two USD files you'd like to compare."

---

Step 1 — Collect metrics from both stages

Write a temp script and run it under the interpreter resolved below. The script collects the same metrics for both stages in one pass.

import json, os, sys
from pxr import Usd, UsdGeom, UsdShade

def safe_getsize(path):
    try:
        return os.path.getsize(path) if os.path.isfile(path) else 0
    except (FileNotFoundError, OSError):
        return 0

def collect(path):
    stage = Usd.Stage.Open(path)
    if not stage:
        return {"error": f"Failed to open: {path}"}

    mpu = UsdGeom.GetStageMetersPerUnit(stage)
    up = UsdGeom.GetStageUpAxis(stage)

    prims = list(stage.TraverseAll())
    meshes = [p for p in prims if p.IsA(UsdGeom.Mesh)]
    materials = [p for p in prims if p.IsA(UsdShade.Material)]
    instances = [p for p in prims if p.IsInstance()]

    total_verts = 0
    total_faces = 0
    for m in meshes:
        mesh = UsdGeom.Mesh(m)
        pts = mesh.GetPointsAttr().Get()
        fvc = mesh.GetFaceVertexCountsAttr().Get()
        if pts:
            total_verts += len(pts)
        if fvc:
            total_faces += len(fvc)

    file_size = safe_getsize(path)

    return {
        "path": path,
        "file_size_bytes": file_size,
        : mpu,
        : (up),
        : (prims),
        : (meshes),
        : (materials),
        : (instances),
        : total_verts,
        : total_faces,
    }

before = collect(sys.argv[])
after = collect(sys.argv[])
(json.dumps({: before, : after}, indent=))

Save the script somewhere writable (the OS temp dir works for ad-hoc runs; use a project scratch dir if you need to keep the artifact). Resolve pxr deterministically — prefer a built repo's bindings (export the build env, mirroring tools/validators/run.sh:20-21 / run.bat:25-26), else use whatever pxr the interpreter already has (wheel/Kit/conda). Don't bare-probe first — a stray PYTHONPATH can shadow the build and fail with a confusing ABI error:

# POSIX
CONFIG="${USD_OPTIMIZE_CONFIG:-release}"; PLATFORM="${USD_OPTIMIZE_PLATFORM:-linux-x86_64}"
BUILD_DIR="_build/$PLATFORM/$CONFIG"; USD_DIR="_build/target-deps/usd/$CONFIG"
if [ -d "$BUILD_DIR" ]; then
    export LD_LIBRARY_PATH="$BUILD_DIR/lib:$BUILD_DIR/extraLibs:$USD_DIR/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
    export PYTHONPATH="$BUILD_DIR/python:$USD_DIR/lib/python${PYTHONPATH:+:$PYTHONPATH}"
    PYBIN="_build/target-deps/python/bin/python3.12"
else PYBIN="python3"; fi
tmpdir="${TMPDIR:-/tmp}"
"$PYBIN" "$tmpdir/_compare_stages.py" "<before.usd>" "<after.usd>"

# Windows (PowerShell)
$Config = if ($env:USD_OPTIMIZE_CONFIG) { $env:USD_OPTIMIZE_CONFIG } else { "release" }
$Platform = if ($env:USD_OPTIMIZE_PLATFORM) { $env:USD_OPTIMIZE_PLATFORM } else { "windows-x86_64" }
$BuildDir = "_build\$Platform\$Config"; $UsdDir = "_build\target-deps\usd\$Config"
if (Test-Path $BuildDir) {
    $env:PATH = "$BuildDir\bin;$BuildDir\lib;$BuildDir\extraLibs;$UsdDir\bin;$UsdDir\lib;$env:PATH"
    $env:PYTHONPATH = "$BuildDir\python;$UsdDir\lib\python;$env:PYTHONPATH"
    $PyBin = "_build\target-deps\python\python.exe"
} else { $PyBin = "python" }
& $PyBin "$env:TEMP\_compare_stages.py" "<before.usd>" "<after.usd>"

For a build-vs-build comparison, use the build's own USD (path 1) so both stages read through the same USD the operations wrote. If you fall back to an installed pxr, pin it to the build: pip install usd-core==25.11 (match the USD version pinned in deps/usd_flavors.json / deps/usd-lib-deps.json; a bare pip install usd-core may not match).

Step 2 — Present the comparison table

Parse the JSON output and present a table with deltas:

Comparison: <before_basename> → <after_basename>

| Metric          |    Before |     After |   Delta |     % |
|-----------------|-----------|-----------|---------|-------|
| File size       |   12.3 MB |    8.1 MB |  -4.2 MB | -34% |
| Prims           |    12,450 |     8,230 |  -4,220 | -34% |
| Meshes          |     3,200 |     1,100 |  -2,100 | -66% |
| Vertices        | 1,245,000 |   620,000 | -625,000 | -50% |
| Faces           |   830,000 |   415,000 | -415,000 | -50% |
| Materials       |        45 |        12 |     -33 | -73% |
| Instances       |         0 |       800 |    +800 |    — |
| metersPerUnit   |      0.01 |      0.01 |       — |    — |
| upAxis          |         Y |         Y |       — |    — |

Format file sizes in human-readable units (KB/MB/GB). Use commas for large numbers. Flag any change in metersPerUnit or upAxis with a warning — those shouldn't change during optimization.

Headline summary

After the table, add a 1–2 sentence synthesis:

Optimization reduced vertex count by 50% and mesh count by 66% (mostly from deduplicateGeometry creating 800 instances). File size dropped 34%. Stage metadata unchanged.

If you know which operations were run (e.g. from a prior run-operations session), attribute the changes to specific ops.

Step 3 — (Optional) Diff validator summaries

When the user passes --validators or asks "did the validator issues improve?", use the summarizer on each stage's saved CSV and diff the totals sections. If saved CSVs don't exist for one or both stages, tell the user to run /run-validators on each stage first. Can check with --verbose flag for running validation for more prim paths.

python3 tools/validators/summarize_csv.py "<before_csv>"
python3 tools/validators/summarize_csv.py "<after_csv>"

Present the per-rule delta:

| Rule                              | Before | After | Delta |
|-----------------------------------|--------|-------|-------|
| UsdOptimizeDuplicateGeometry   |    120 |     0 |  -120 |
| UsdOptimizeColocatedVertices   |     45 |     0 |   -45 |
| ...                               |        |       |       |

Rules that dropped to 0 are resolved. Rules that didn't change may need different operations or parameter tuning.

Step 4 — (Optional) Prim-level diff

When the user passes --prims or asks "which prims were added/removed?", add this helper to the Step 1 script. It reopens the two input paths instead of trying to reuse collect()'s local stage variable:

def collect_prim_paths(path):
    stage = Usd.Stage.Open(path)
    if not stage:
        return set()
    return {str(p.GetPath()) for p in stage.TraverseAll()}

before_paths = collect_prim_paths(sys.argv[1])
after_paths = collect_prim_paths(sys.argv[2])

added = sorted(after_paths - before_paths)
removed = sorted(before_paths - after_paths)

For large stages this can produce thousands of lines. Default to showing the first 20 adds and 20 removes, with a count of how many more exist. Expand on request.

Prims that exist in both stages but changed (e.g. vertex count dropped) require per-prim attribute comparison — only do this if the user asks about a specific prim path.

---

See also

• .agents/skills/run-operations/SKILL.md — producing the optimized stage.
• .agents/skills/run-validators/SKILL.md — generating validator artifacts.
• .agents/skills/inspect-asset/SKILL.md — single-stage inspection.
• .agents/skills/config-presets/SKILL.md — what each preset config is expected to change.

Purpose

Quantify what an optimization (or any stage edit) changed. Produces a structured before/after report at three levels — file size, geometry counts (prims, meshes, vertices, faces, materials, instances), and optionally validator-finding deltas. Use when the user asks "what changed?", wants to confirm a pipeline did the expected work, or needs to attribute a regression to a specific operation.

Prerequisites

• Two USD files (.usd / .usda / .usdc / .usdz) on disk.
• A Python interpreter with the USD bindings (from pxr import Usd) — a built repo (_build/target-deps/, env exported per Step 1), an existing pxr (wheel/Kit/conda), or a pinned pip install usd-core==25.11.
• For --validators: saved issues.csv artifacts from run-validators for both stages.

Limitations

• The skill is read-only — it never modifies either input.
• Per-prim attribute diffing (e.g. "did vertex positions change?") is out of scope; ask the user to point at a specific prim if they want attribute-level comparison.
• Validator diff requires both stages to have been run through the same validator set (mismatched rule sets produce noisy deltas).
• Prim-level diff (--prims) can be slow and verbose on multi-million-prim stages — defaults to first 20 adds/removes.

Troubleshooting

Symptom	Likely cause	Fix
Failed to open: <path> in the JSON output	The path doesn't exist or isn't a valid USD layer.	Verify the path; check the file extension is .usd*.
Validator diff says "no saved artifacts"	One or both stages haven't been validated yet.	Run run-validators on each stage first to produce issues.csv.
Counts identical despite an edit	Edit was authored on a sublayer the open call doesn't see.	Open the asset with the same layer-stack / session that produced the edit.
metersPerUnit or upAxis flagged as changed	An operation rewrote stage metadata it shouldn't have.	Check the pipeline — most Usd Optimize ops never touch stage metadata. Surface as a warning.
Prim-diff returns thousands of entries	Edit was a wholesale tree rebuild (e.g. flatten).	Switch to summary-level reporting; ask the user if they want a specific subtree compared.