رفتن به محتوا
newblender

14 — The Junk Shop (Blender 2.81 splash): how an agent could make it

این محتوا هنوز به زبان شما در دسترس نیست.
Original Alex Treviño (concept by Anaïs Maamar), 2019, stylised interior environment, Blender 2.81 splash · BlenderNation · blender.org demo files
Agent-readiness today Taste-bound: every mechanic (props as data, declared modifiers, set dressing by raycast, Cycles lighting) works through the live bridge today. What made this scene a landmark is a concept artist’s design, hundreds of hand-shaped stylised props with personality, painted texture wear and a lighting mood. Those are judgement, and hand texture painting is also stroke-based (L-009).
Difficulty 4 (a solo artist’s months-long piece; an agent can block it in days, but closing the gap to the original’s charm is the long tail)
First slice A shop corner: a 2 m shelf unit and a counter built as data, filled with about 40 props from 6 parametric prop recipes (crates, jars, bottles, book stacks, a clock, a lamp) dropped onto their shelves by raycast, Poly Haven CC0 materials, a warm practical lamp against a cool window, rendered in Cycles.

1. What the original actually is

Section titled “1. What the original actually is”
  • One still image (the 2.81 splash) and one public .blend: blender.org’s demo-files page lists “Blender 2.81 – The Junk Shop”, Alex Treviño, original concept by Anaïs Maamar (verified 2026-10-07; this settles the index’s unverified concept credit).
  • Content (from the image, interpretation): one interior room seen from a single camera; a counter, shelves and floor piled with props (radios, lamps, frames, boxes, bottles, a bicycle-scale piece of machinery), warm interior practicals, light from a window, stylised proportions and painted wear.
  • Afterlife: a standard benchmark scene. Phoronix’s pts/blender test added the Junkshop scene in version 4.1.0 (March 2024) and it is one of its most-run scenes (verified via search); it is believed to be one of the Blender Open Data benchmark scenes (unverified). It has also been ported to other renderers (e.g. a Guerilla Render version), which says the file is treated as a reference scene.

2. How the humans made it

Section titled “2. How the humans made it”

Stages: S01 Modeling, S02 UV & texturing, S03 Shading, S06 Layout, S09 Lighting & rendering.

  • Fact (index, BlenderNation): solo artist; prop modelling, UV and texture work, layout and set dressing, Cycles lighting. The article itself gives no process detail; the artist’s Blender Artists thread is where process notes would be (not fetched).
  • Interpretation (typical for this kind of scene): each prop is box-modelled with Subdivision and Bevel, shapes exaggerated for style; textures mix procedural grime with painted masks; set dressing is by hand in the viewport, prop by prop, judged through the camera; lighting is a small number of warm practicals plus a cool key, tuned by eye.
  • Interpretation: the S01 census pattern holds here too: a low-poly cage plus live modifiers is the dominant authored artefact (91% of asset meshes carry modifiers; Subdivision 54% of all modifiers; median mesh 111 vertices; docs/discovery/s01-modeling/02-census.md).

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 (free sources) Tooling
S01 Room and furniture Walls, floor, shelf unit, counter as bmesh data with real dimensions; Bevel and Subdivision declared, not applied (chair-demo method) Real furniture dimensions tools/live/bl.py, bmesh → Mesh
S01 Props ~6–10 parametric prop recipes as GN groups or Python builders (crate, jar, bottle via revolve, book stack, clock, lamp, radio box); each prop a seed + parameters, stored in a JSON prop manifest None GN; data builders
S01 Props (hero) Imported CC0 models where a recipe would be poor (old radio, tools, chairs) Poly Haven models (CC0), Kenney (CC0, stylised low-poly) glTF import
S02 UVs Recipes emit UVs as data (the chair demo learnt that calc_uvs needs an existing layer); imported assets keep theirs None bmesh
S03 Shading Library materials: worn wood, rusty metal, glass, cardboard, plus a GN-written wear attribute (edge curvature, height) feeding a dirt mix Poly Haven textures, ambientCG (CC0) Shader Attribute, Geometry > Pointiness in Cycles
S06 Set dressing Placement as data (L-009): for each manifest entry, raycast down onto a declared support surface, random yaw within limits, reject on BVH overlap; “clutter density” per shelf as an input None mathutils.bvhtree
S09 Lighting One cool area light at the window, 2–3 warm point/area practicals inside lamp props, low world fill; camera from the concept (owner gives framing) Optional Poly Haven HDRI (CC0) seen through the window Cycles, fixed seed

Build order:

  1. Owner supplies or approves a framing and a mood reference (the 2.81 splash itself is the obvious one).
  2. Write checks (§5).
  3. Room shell and furniture as data; camera locked to the framing.
  4. Prop recipes one by one; contact sheet of seeds for owner review.
  5. Prop manifest (counts per shelf), then raycast set dressing; overlap and floating checks.
  6. Materials and wear attribute.
  7. Lighting pass; test renders at 25%; owner review; full render.

4. What works today vs. what needs newblender

Section titled “4. What works today vs. what needs newblender”
  • Works today with the live bridge:
    • Geometry as data with declared modifiers, validity and spec checks: proven by demo 001 (12/12 checks, ~/newblender-data/demos/001-chair/report.md).
    • Raycast and BVH overlap tests for placement; reading evaluated geometry.
    • Importing glTF/FBX assets; assigning library materials; Cycles render to file; viewport screenshots for the owner.
  • Painful today:
    • Set dressing in stock Blender assumes a mouse: drag, snap, rotate in the viewport. Placement has to be re-expressed as data (L-009), which we do in our own code; the editor gives no “place on surface” function with explicit inputs (L-004: snapping reads hidden tool settings).
    • Hand texture painting is a brush stroke stream with no data form (L-009); we substitute procedural wear masks, which look procedural.
    • Iterative look tweaks from Python push no undo (L-008); the owner cannot step back through agent changes.
    • No machine-readable result for a placement batch (what moved, what collided): our checker reconstructs it (L-003).
  • Not feasible today:
    • Props with the original’s stylised character (exaggerated bends, hand-sculpted dents) from recipes alone. Recipes give tidy, plausible props; charm needs either a human pass or an image-to-3D model whose output we would then have to clean.
    • Painted texture storytelling (labels, stickers, specific scuffs) at the original’s density.

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

Section titled “5. Verification plan (RFC 0001 R6, six layers)”
Layer Check Threshold
1 Validity All meshes clean after modifiers 0 non-manifold edges (closed props), 0 degenerate faces, 0 loose verts, normals outward
1 Validity Every texture path resolves 0 missing images
2 Spec Prop count and spread ≥ 40 props in the slice; every declared shelf holds its manifest count ± 1
2 Spec Physical set dressing Every prop’s lowest point within 2 mm of its support; 0 pairwise BVH overlaps
2 Spec Hygiene Scale applied (1.0) on all objects; Studio naming GEO-<prop>_<n>; UVs on every mesh; material on every face
2 Spec Budget Slice < 1.5 M faces; 1920 × 1080 Cycles render < 15 min on the owner’s GPU
3 Reference Statistical comparison with the public Junk Shop .blend (fetch later from blender.org demo files; not downloaded now): props per visible m², light count, materials count, faces per prop, camera focal length Report deltas; gate only on light count within ±2 and focal length within ±10 mm of the reference
4 Downstream Render health 0 NaN; fireflies < 0.01% pixels > 20× local median; noise after denoise: SSIM vs a 4× sample render ≥ 0.97
4 Downstream Scene opens headless and renders with the same result Headless vs live render PSNR ≥ 45 dB
5 Appearance (warning) Vision model compares the render with the 2.81 splash on “cluttered, warm, cosy junk shop” and lists missing prop categories Warn below 6/10
6 Taste (owner) Owner judges charm, clutter rhythm and light mood Sign-off; notes become manifest and light parameter changes

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • The concept. The original had a concept artist before a 3D artist; someone must decide what this shop is, who owns it, what story the clutter tells.
  • Prop personality. Exaggeration, asymmetry and wear that read as “loved objects” are a modeller’s touch; the agent’s recipes are the blocking pass.
  • Composition through clutter. Leading lines and quiet zones in a busy frame are placed with intent; random-with-limits fills space evenly, which reads as noise.
  • Light mood. The balance of warm practicals against cool daylight is tuned by eye, usually over many small iterations.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, ~1 day): in the owner’s Blender window, an empty room becomes a wall corner with a wooden shelf unit and a counter. Then shelves fill: jars, bottles, crates, book stacks, a clock and a lamp drop into place one shelf at a time, each resting on its board, none overlapping. Changing a shelf’s density input re-dresses that shelf. The viewport switches to rendered mode: warm lamp glow on the counter, cool window light on the back wall. A Cycles still and a check report (counts, contact, overlaps, hygiene) land in ~/newblender-data/demos/.
  • Full reproduction: 80–150 agent-hours for room, 30+ prop recipes, set dressing and checks; 1–3 GPU-hours of final renders; 40–100 human-hours of concept, prop polish and lighting direction to approach the original’s quality.

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • Recipe props may look generic; the gap to the original could be mostly in the long tail of hand detail.
  • The demo-files link goes to the old Blender Cloud gallery; the file may now be hosted on Blender Studio. Licence of the .blend is not stated on the listing (check before reuse).
  • Vision-model scoring of “cosy” is unreliable (R6 layer 5 is a warning only).
  • Image-to-3D models could raise prop quality but bring licence and topology questions (L-010 QA gates would catch topology, not licence).

9. Sources

Section titled “9. Sources”