Bake Normal High To Low

A runnable example that cage-bakes a ribbed bronze hatch plate onto a DECIMATE COLLAPSE LOD and asserts the tangent-space normal map carries measurable surface detail — following bake-high-to-low.

What it witnesses: Cycles selected-to-active normal bake is a statistical process, not a byte-identical one. A high-poly source produces a map that deviates from flat tangent (0.5, 0.5, 1.0); the same bake from an undisplaced source does not.

Byte-identity across 4.5 / 5.1 / 5.2 is not the contract. Tile order and float accumulation differ even at one CPU sample. The gates are fraction of pixels beyond Euclidean 0.04 from flat, mean absolute deviation, and a monotonic gap versus a flat control. Tolerances sit well inside the measured gap (detail frac 0.7211 vs flat 0.0000) so they are not tuned-until-green.

• Detail bake is not flat. frac >= 0.40 and MAD >= 0.05 (measured 0.7211 / 0.09356 on 4.5.11, 5.1.2, and 5.2.1). Catches an inactive Image Texture node, reversed selection, or EEVEE/GPU mis-setup that writes a blank map. MAD floor is half the measured hatch value, still ~30× a flat bake.
• Flat control is flat. frac <= 0.05 and MAD <= 0.03 (measured 0.0000 / 0.00277). Catches a noisy or wrongly-typed bake that would also satisfy the detail gates.
• Monotonic gap. detail_frac - flat_frac >= 0.30. The two maps must separate; a tolerance wide enough to pass both would have no discriminating power.
• --flat-source is the falsifier. Skips the ribs and still runs the detail gates. Must exit 5. Analogous to --same-axis in export-preset-axis.
• The target node must be selected (5.0 and later). The bake writes into the material's active Image Texture node, and on 5.x only if that node is also selected. --unselect-target keeps nodes.active = tex but sets tex.select = False. Measured: 5.0.1, 5.1.2 and 5.2.1 return {'CANCELLED'} with Info "No active and selected image texture node found" and leave the pixels untouched, so the run exits 4. On 4.5.11 the bake ignores selection, finishes, and the run exits 0. The catalog falsifier therefore carries "min_version": "5.0" and is recorded as SKIP on 4.5.

Neighbor of lod-decimate-chain (the LOD is the cage target; collapse keeps UVs) and image-pixels-testcard (save_render, not Image.save(), if you persist the datablock). UV transfer and atlas packing are out of scope.

The still reads left to right as the pipeline: the baked map as an unlit card, the high-poly source it was baked from, and the collapse-decimated LOD wearing it. Source and LOD share one cast-bronze material, so the only difference between them is real geometry versus the map. The floor captions carry live triangle counts from the evaluated meshes (3200 vs 900). A raking key from low on the left makes both the real relief and the baked relief throw light and shade. If the bake were flat, the card would be uniform (128, 128, 255) periwinkle and the right plate would shade like the undisplaced cage while the middle one kept its ribs.

Staging is render-only. Each panel stands in a low display plinth, lifted so its lowest evaluated point sits 35 mm in the slot. Both panels used to balance on an edge, and the leaning plate pierced the floor. The LOD's solidify rim wears plain bronze instead of the baked map. The slight waviness along the plate's border is the bake itself, normal-map shading inside the UV margin, and is left as the map produces it. Bake statistics are unchanged: frac 0.7211, MAD 0.09356 on all three binaries.

Operator RNA (type='NORMAL', use_selected_to_active, cage_extrusion, cage_object as a string, normal_space='TANGENT', margin_type) is identical on 4.5.11, 5.1.2, and 5.2.1 — no shim.

Run#

# Cheap correctness check (no render) - the CI check:
blender --background --python bake_normal_high_to_low.py --

# Falsifier: undisplaced high. Must exit non-zero (detail frac gate).
blender --background --python bake_normal_high_to_low.py -- --flat-source

# Falsifier: active but deselected target node. Exits 4 on 5.x, 0 on 4.5 LTS.
blender --background --python bake_normal_high_to_low.py -- --unselect-target

# Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
blender --background --python bake_normal_high_to_low.py -- --output hatch.png
blender --background --python bake_normal_high_to_low.py -- --output hatch.png --engine cycles

Exit codes#

Per-script sequential checks. 9 is a valid check code; there is no rule against it. 10 is the shared framing helper.

The blender-smoke workflow runs the check on Blender 5.2 LTS and 4.5 LTS (5.1 on the weekly cron, the needs-5.1 PR label, or manual dispatch). Smoke does not pass --output.