Claude Opus 5.5 rebuilds a scanned garden in Blender with Geometry Nodes

• A garden splat became an editable Blender scene: gokayfem's demo used Claude Opus 5.5 for procedural objects, Geometry Nodes, Shader Nodes and captured textures.
• The creator reports no manual edits: he said the process was automatic.
• The workflow is open source: gokayfem shared the repository with an agent skill, reusable prompts and reconstruction tools.
• Fidelity remains approximate: the creator described the result as “not insanely accurate”.

Gokayfem said Claude figured out more than scale; the Garden notes anchor scale to an assumed 74-centimeter table height. The release provenance contains another wrinkle: the finished video's table and paving are baked mesh copies carrying captured textures.

The reference is the Mip-NeRF 360 Garden scene, displayed through GaussianSplats3D, according to the project documentation.

The published Garden builder divides the environment into seven parts:

• Ground: slab layout, bevels, joints, litter, soil and lawn.
• Table: radial planks, slate center, supports and materials.
• Vase: asymmetric vessel, opening and dried botanical arrangement.
• Tree: trunk, branches, forks, bark and canopy.
• Shrubs: density controls, branches and leaf instances.
• Boundary: garden edges and surrounding geometry.
• Structures: beds, pots, soil, stakes, seating and props.

Semantic parameters live in JSON files; the builders produce Geometry Node groups and Shader Node materials. That editable object breakdown is the prize here.

The reusable prompt pack documents the acquisition, calibration and correction work behind the reconstruction:

1. Project brief: establishes the reference, whole-scene scope, environment and fidelity target.
2. Capture and registration: produces repeatable views, shared measurements and matched cameras before detailed modeling.
3. Part contracts: assigns each substantial component shared inputs, ownership and verification views.
4. Detail and review: follows an integrated blockout with repeated visual correction.
5. Hybrid projection: optionally maps captured appearance onto reconstructed surfaces.
6. Comparison and delivery: packages source, structural checks, matched views and optional comparison video.

“Capture” means rendering and measuring an existing splat. The templates generalize the Garden process and were rewritten for reuse; they are not a verbatim session transcript.

Gokayfem said he did not touch the result manually. In a workflow reply, he listed three assignments:

• Opus 5.5 Medium: subagents.
• Opus High: orchestrator agent.
• Astra Extra High: 10-minute visual-feedback passes.

The released prompt pack also supports a single agent. Delegation is optional.

The mapping workflow takes an existing reconstruction and calibrated captures through three stages:

1. Prepare: Blender generates evaluated UV surfaces and depth/object visibility maps.
2. Project: a NumPy/SciPy/Pillow pipeline projects captured images onto those surfaces, applying visibility and rejection rules.
3. Assemble: Blender packs the results onto selected mesh copies. Unseen or rejected texels retain the original material.

Each project supplies its own objects, material slots, cameras, masks and projection settings. Objects visible only in the source captures require masks or other explicit evidence, because the reconstructed scene's depth cannot identify them.

The original procedural source remains separate, and untouched parts retain their live node graphs. Geometry or UV edits to mapped surfaces require rebaking.

The two appearance modes offer unlit captured radiance or an approximate lighting-compensated material. Neither recovers clean physical albedo; the release retains baked illumination, some stretched appearance and coverage gaps.

The native-scene audit removes the reference in memory and compares mesh and instance signatures. It tests whether the reconstructed geometry survives independently of the scan.

Visual comparison uses paired reference and Blender frames from shared camera manifests, with hashes recorded for the evidence. The release's comparison turn uses 288 shared poses at 1.25-degree increments.

The Garden study documents remaining differences in:

• Foliage distribution and plant morphology.
• Inferred hidden surfaces.
• Table and vase construction.
• Paving wear, materials and lighting response.

Gokayfem also welcomed a suggestion for improving positional accuracy. His own assessment stayed modest:

The installation documentation specifies Python 3.10+ and Blender 5.0+, with local integration verification on Blender 5.0.1. The documented Agent Skills installation command is:

Claude Code plugin installation and manual Codex installation are also documented. A skill-only install includes the tools and a synthetic editable-table starter; the full repository contains the Garden study.

The video records a separately fitted hybrid experiment. Its fitted Garden textures, capture arrays and binary .blend files are excluded from the repository, along with the original scan.

MIT covers the repository's code and instructions. Dataset assets and third-party media retain their own terms.