رفتن به محتوا
newblender

03 — Charge: how an agent could make it

این محتوا هنوز به زبان شما در دسترس نیست.
Original Blender Studio (dir. Hjalti Hjálmarsson, art director Andy Goralczyk, lead animator Rik Schutte), 2022, open movie: photoreal action · Blender Studio — Charge
Agent-readiness today Partial. The hard-surface factory, the PBR materials, the GN smoke, fire and muzzle-flash FX, and the lighting and rendering are agent-buildable today. The realistic digital human (Einar), his hair groom, facial performance and combat animation are not.
Difficulty 5 (only about 3 minutes, but a believable digital human in heavy action is among the hardest things in CG)
First slice A battery-factory bay built as a modular kit: arrayed shelving racks of glowing battery cells, declared-bevel walls and pipes, CC0 PBR metal and concrete, volumetric haze and a GN muzzle-flash and smoke burst, rendered as a Cycles still plus a short EEVEE camera push-in.

1. What the original actually is

Section titled “1. What the original actually is”
  • Film: “a high-visual-impact, action-packed 3-minutes-long animation inspired by the game cinematics and realtime demos formats” (Blender Studio, verified 2026-10-07).
  • Hero assets:
    • Einar: a realistic old man, sculpted in layers, with a curves-hair groom and facial rig.
    • The security robot: many modelling iterations.
    • Packbot: a smaller robot with driving tests.
    • The battery factory: a hard-surface environment.
  • FX: smoke, fire and muzzle flashes made with Geometry Nodes.
  • Output: rendered with both EEVEE and Cycles comparisons (“EEVEE and Cycles rendering comparisons”, same page).
  • Released: the “Charge Material Asset Library”, as a freebie. Shot files are for subscribers.
  • Shots: the count is not published. For a 3-minute action short, roughly 40–80 is my estimate (interpretation).

2. How the humans made it

Section titled “2. How the humans made it”

Facts:

  • Stages: S01, S02, S03, S04, S05, S07, S08, S09, S10 (index §03).
  • Einar was built with layered sculpting and the then-new curves hair system (index; Layered sculpting for Einar, unverified).
  • Geometry Nodes for smoke and fire, curve-stitching tools for modelling, and advanced facial rigging (Blender Studio, verified).
  • The project began as “Project Heist” and pursued realism through interactive PBR workflows (same page; index).

Interpretation: Charge splits cleanly in two.

  • The world: factory, robots, materials, lighting and FX. This is engineering-shaped: dimensions, modular repetition, physically based values and procedural FX. It is the agent’s natural territory.
  • The human: skin, hair, micro-expression and weight in a fight. This is the uncanny-valley problem, where small errors are fatal and judgement is entirely perceptual.

3. How I would make it: the agent-native plan

Section titled “3. How I would make it: the agent-native plan”
Stage Agent approach Inputs Tooling
S01 Modeling (environment) A modular kit as data: wall panel, floor tile, pillar, pipe run and battery rack. Each part is a low cage with declared Bevel (harden normals), Array, Mirror and Solidify. Racks are a GN instancer over a grid with per-instance variation. Real dimensions for industrial racking (e.g. 2.0 m bays, 0.6 m deep); reference photos bl.py, bmesh-as-data, L-006
S01 Modeling (droid) Blockout as declared primitives plus Booleans, then a support-loop pass. The final hero droid needs a human design pass. Concept art (human) Declared Boolean, Bevel and Weighted Normal stack
S01 Modeling (Einar) Not attempted. Use a CC-BY base mesh only as a stand-in. human-base-meshes-bundle-v1.4.1 (local) —
S02 UV & texturing Box or cube projection on the kit; trim-sheet UVs on panels; numeric stretch checks. Poly Haven CC0 metal, concrete, painted-steel and rust textures; the Charge material library (download) UV by script
S03 Shading PBR from measured values: metalness 0/1, roughness bands, emissive battery cells with a declared colour temperature. Charge Material Asset Library (CC-BY freebie) Shader nodes and asset linking
S04 / S08 FX Muzzle flash as GN instanced emissive cards with a 2–3 frame lifetime. Smoke as a Mantaflow gas burst, or GN points to Volume with animated noise density. Sparks as GN points with gravity. None GN simulation zone; Mantaflow -b bake
S05 Rigging Droid: a mechanical rig of bones parented to parts, with joint limits as data (no skin weights needed). — Armature by script
S07 Animation Droid patrol paths and mechanical cycles are scriptable. Einar’s performance and the fight choreography are human. Previs (human) F-curves by script
S09 Lighting & rendering Practical lights from the declared rack emitters, cold key and warm rim, volumetric haze; Cycles finals and an EEVEE preview. — Headless renders
S10 Compositing Glare on the emitters, lens distortion and grain as a declared compositor graph. — Compositor nodes by script

Build order (for the first slice):

  1. Write the checks first: bay dimensions, rack count, PBR value ranges, haze and emission levels, FX timing.
  2. Set up a clean slate with metric units. Build the kit parts as data, each with its declared modifier stack, and check validity per part.
  3. Lay out a 20 × 12 × 6 m bay: two rows of racks via GN, pillars arrayed at 6 m, and pipe runs along the ceiling.
  4. Import Poly Haven materials (or link the Charge library) and assign them by naming rule (GEO-rack_* gets painted steel, and so on).
  5. Battery cells: an emissive shader with per-instance random flicker driven by a GN attribute.
  6. Add a volume cube with haze density, a key light through a ceiling grille, and the emitters as practicals.
  7. FX beat at frame 48: a GN muzzle flash (2 frames), a spark burst, and a smoke puff.
  8. A camera push-in of 72 frames. Show progress in the window with snap.py, render the Cycles still headless at 1920 × 1080, and render an EEVEE preview of the 72 frames.

4. What works today vs. what needs newblender

Section titled “4. What works today vs. what needs newblender”
  • Works today with the live bridge:
    • Kit parts as bmesh data with declared stacks (proven by the chair demo).
    • GN instancing and simulation zones built by script.
    • Material import and assignment.
    • Volumes, light set-up, the compositor graph, and headless Cycles and EEVEE.
  • Painful today:
    • Hard-surface refinement (support loops, knife cuts, inset panel lines) is local topology surgery with no declarative twin. 59 such operators need edit mode plus selection (05-brief §4; L-001, L-002, L-005). Declared Bevel and Boolean cover most of it, but not all.
    • Boolean cleanup: the results are hard to validate without a returned diff (L-003). The validity check catches non-manifold output after the fact.
    • Mantaflow bakes are operator-only with context polls (L-004).
    • Hair grooming (curves) is brush-driven: comb, cut and puff strokes (L-009). That is a human act with no data API for “comb this way”.
  • Not feasible today:
    • A believable realistic human: skin shading tuned by eye, facial performance, hair groom.
    • Fight choreography with weight and intent.
    • The robot’s character design.

5. Verification plan (RFC 0001 R6, six layers)

Section titled “5. Verification plan (RFC 0001 R6, six layers)”
Layer Check Threshold
1 Validity Each kit part, evaluated: non-manifold edges, degenerate faces, flipped normals, loose verts 0 / 0 / 0 / 0
1 Validity Boolean results: self-intersecting faces (BVH overlap test) 0
2 Spec Bay bounding box 20 × 12 × 6 m ± 2%
2 Spec Rack count / bay spacing 2 rows × 8 racks / 2.0 m ± 1 cm
2 Spec Face budget of the whole bay, evaluated ≤ 1.5 M faces
2 Spec PBR sanity: every metal has metallic = 1 and roughness 0.15–0.7; every dielectric has base-colour luminance 0.03–0.9 100% of materials
2 Spec Muzzle flash visible frames / smoke lifetime 2–3 frames / 24–48 frames
2 Spec Applied scale and naming (GEO-, FX-, LGT-) all 1.0 / 100%
3 Reference Rebuild from the recipe in a fresh headless session: evaluated mesh hash per kit part identical
4 Downstream UV stretch (area-distortion ratio) on textured parts ≤ 1.15 for 95% of faces
4 Downstream Cycles 1080p still at 256 samples with denoise: time, NaN or fireflies (pixels > 50 × the median luminance) < 15 min; 0; < 20
4 Downstream EEVEE preview of 72 frames: seconds per frame ≤ 5 s
5 Appearance (warning) Vision model compares to Charge stills: an industrial, dim, cold-key, photoreal bay warning if under 3/5
6 Taste (owner) Does it read as Charge’s world (oppressive, game-cinematic, real)? owner’s call

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • Einar end to end: the likeness, sculpt detail, skin shading, groom and facial rig. Realism means perceptual judgement in every iteration.
  • Action choreography and animation: weight, impact, and the old man’s desperation.
  • Robot design: the silhouette and the threat it reads as.
  • Art direction: Andy Goralczyk’s palette, the haze level, where the light falls in each shot.
  • The cinematic edit: pacing a 3-minute action piece.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, about 1 day):
    • On screen: in the owner’s window, an empty grid fills with a concrete floor. Then two long rows of steel shelving racks appear and multiply down the hall as the Array and GN instancers take effect. Ceiling pipes and pillars follow.
    • Hundreds of battery cells light up cyan with a slight random flicker. Fog fills the hall, and light shafts come through a ceiling grille.
    • Scrubbing to frame 48 shows a bright muzzle flash, a shower of sparks and a smoke puff.
    • The final Cycles still and a short EEVEE push-in clip render headless, with the check report.
  • Full reproduction:
    • Agent work: about 200–400 agent-hours for the environment, droid modelling support, FX, lighting and render.
    • Render compute: about 4,300 frames × 5–20 min Cycles ≈ 360–1,400 GPU-hours.
    • Human work: the Einar package alone is months of specialist time (sculpt, groom, shading, face rig), plus animation of about 3 minutes of action by a small team. Overall about 1 studio-year, as the original took.

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • The Charge material library licence and download format are not verified here. Fall back to Poly Haven CC0 if needed.
  • The GN smoke look versus Mantaflow: which reads as “real” at the Charge level is untested.
  • EEVEE Next and Cycles parity for emissive-heavy, volumetric scenes may differ from 2022 behaviour.
  • No Charge shot files are on this machine (they are subscriber-only), so layer-3 reference checks against the original are not possible; only self-consistency.

9. Sources

Section titled “9. Sources”