Skip to content
newblender

Demo 003: Charge, first slice — report

Battery-factory bay from docs/showcase/03-charge.md, built live in Blender 5.2.2 through the Blender Lab bridge. Build: tools/live/demos/charge_build.py. Checks, written before the build: tools/live/checks/charge_checks.py. Later scripts: charge_l3_l4.py, charge_render_still.py. Textures: Poly Haven CC0 (concrete_floor_worn_001, concrete_floor_02, blue_metal_plate, metal_plate, corrugated_iron), 1K, in ~/newblender-data/assets/polyhaven/.

Outputs: charge.blend · still.png / still.exr (Cycles, 1080p, frame 56) · charge_pushin_eevee_1080p.mp4 (96 frames) · eevee_frames/

Results

Section titled “Results”
Result Layer Check Measured
PASS 1 kit parts: non-manifold / degenerate / inward / loose 0 in 9 parts
PASS 1 boolean wall: self-intersecting face pairs 0
PASS 2 bay bounding box 20 x 12 x 6 m +-2% 20.00 x 12.00 x 6.00
PASS 2 racks: 2 rows x 8 [8, 8]
PASS 2 rack pitch 2.0 m +- 1 cm 2.000-2.000
PASS 2 face budget <= 1.5 M 20,295
PASS 2 PBR sanity (metal roughness 0.15-0.7; dielectric luminance 0.03-0.9) all ok
PASS 2 muzzle flash visible frames (2-3) [48, 49]
PASS 2 smoke lifetime 24-48 frames 42
PASS 2 scale applied all 1.0
PASS 2 naming GEO-/FX-/LGT-/CAM- 100%
PASS 3 fresh headless rebuild gives an identical fingerprint 2c844a062d01ce49 in both (768 cell instances)
FAIL 4 UV stretch: ≥ 95% of faces within 1.15 area-distortion 12.4% of faces. By surface area: rack 95.9%, ceiling 95.6%, back wall 99.2%, pillars 94.2%. Unbevelled parts are 100%. The failing faces are 8–15 mm Bevel chamfers whose UVs interpolate between two box projections.
PASS 4 Cycles still, 1080p, 256 samples + denoise, < 15 min 13.0 min on Apple M4 Max (GPU - 32 cores)
PASS 4 NaN/Inf pixels 0
FAIL 4 fireflies < 20 (pixels > 50× the median of their 5×5 neighbourhood) 24. All 24 sit in the spark burst, with RGB 1 : 0.53 : 0.19 = the spark emission (1 : 0.55 : 0.2). They are spark streaks, not render noise, but the rule as written counts them.
PASS 4 EEVEE preview, ≤ 5 s per frame median 1.04 s, max 2.78 s (96 frames, 1080p)
WARN 5 agent’s own look vs. Charge (warn under 3/5) — self 2.5/5 (blind: 2.33, see below). Reads as a real industrial aisle: textured painted-steel racks, concrete, steel pillars, slatted ceiling, spark burst. Lighting is too even and bright for Charge’s dark, high-contrast look; cells read as plastic tubs, not glowing batteries; smoke is faint in Cycles and missing in the EEVEE preview.
— 6 does it read as Charge’s world? owner’s call

Fixes along the way

Section titled “Fixes along the way”
  • Look:
    • Cells blew out to white (emission 3.6–6): lowered to 0.65–1.35 with saturated cyan.
    • Muzzle flash was a tiny star: made 2.5× larger.
    • Smoke was invisible: density raised about 4×, radius larger.
  • Smoke in EEVEE: still invisible after the fix. A Cycles render of the same frame shows it, so the shader is correct. This is an EEVEE limitation with a volume object nested in the haze volume.
  • Haze: milky in Cycles, so the density was halved (0.035 → 0.018).
  • PBR (first run 10/11): corrugated iron roughness averaged 0.702 against a 0.70 limit. The fix was an art adjustment (× 0.9 roughness node, now about 0.63). The checker was also corrected: it sampled the raw texture and ignored nodes between texture and shader, so it now applies multiply-by-constant nodes on that path. Both changes are listed here because one touched the checker.

Deviations from the plan

Section titled “Deviations from the plan”
  • Frame range: 96 frames, not 72, so the 42-frame smoke fits inside the range. The push-in spans all 96.
  • Firefly rule: uses a local 5×5 median instead of the global median, so a bright light source isn’t counted. Even so, the sparks count.
  • No compositor pass (glare, lens distortion, grain), no droid and no character. The plan’s first slice didn’t require them.
  • Cycles device: the live session renders Cycles on the CPU (about 37 s at 32 samples, half resolution), so the final still was rendered headless on the GPU. GPU was enabled only inside that factory-startup process; the owner’s preferences were not changed.

For the owner to decide

Section titled “For the owner to decide”
  1. UV metric: should UV stretch be area-weighted? By face count, tiny bevel chamfers dominate.
  2. Firefly rule: should known emitters (sparks) be masked out?

Blind quality (RFC 0001 R7)

Section titled “Blind quality (RFC 0001 R7)”

Session s01-flow-charge, scored by the owner on 2026-10-08 without knowing sources (12 items, 540 px wide). Full table: ~/newblender-data/blind/sessions/s01-flow-charge/results.md (local only).

Agent (Charge slice) Studio (Charge)
Items 3 (Cycles f56, EEVEE f48, EEVEE f80) 2 (film frame, production scene 050_0160 rendered locally)
Mean quality (1–5) 2.33 (2, 3, 2) 4.50 (4, 5)
Taken for the other source 1 of 3 (EEVEE f48, muzzle flash, rated 3, guessed studio) 0 of 2
  • Quality measure: 2.33 / 5. It’s the better of the two slices, and the only agent item in the session that passed as studio work.
  • Self-score calibration: the self score was 2.5, so the agent was close (0.17 over).
  • Telling detail: the frame that fooled the owner is the muzzle-flash frame. Strong light and motion hide the even lighting and plastic-looking cells that give the other frames away. That points to lighting contrast as the next quality lever, which matches the self review.