lumera-single-image-scene-reconstruction

Reconstructs editable Godot 4 scenes with per-object meshes and parametric lights from a single reference image.

66|4|Updated Feb 26, 2026
One-click install
npx skills add https://github.com/SummerEngine/summer --skill lumera-single-image-scene-reconstruction-summerengine
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: lumera-single-image-scene-reconstruction
Source: https://github.com/SummerEngine/summer/tree/main/library/skills/lumera-single-image-scene-reconstruction
Command: npx skills add https://github.com/SummerEngine/summer --skill lumera-single-image-scene-reconstruction-summerengine

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Turning a single concept image or screenshot into a 3D scene usually produces a fused point cloud, Gaussian splat, or baked texture that artists cannot edit. This Skill adapts the Lumera pipeline (arXiv:2607.20889) to SummerEngine/Godot 4 so the output is an engine-native .tscn scene with separate MeshInstance3D objects, countable parametric light nodes, and an HDR environment probe that can be moved, replaced, and re-tuned afterward. ## Core Features & Use Cases - Structured scene parsing: Defines a serializable schema for VLM-parsed object boxes (position, size, yaw) and parametric light tuples (position, color, intensity), with a repair layer for invalid sizes, duplicate IDs, and bad labels. - Engine-native assembly: Generates a text .tscn node tree (Camera3D, WorldEnvironment with HDR sky, Objects and Lights containers) that stays diffable, versionable, and editable in the editor. - Bounded dual-agent refinement loop: Implements a geometry stage then a light stage with sliding-window history, verifier feedback against the reference image, and hard round limits so the loop terminates predictably. - Use Case: Block out a level from one piece of concept art, then hand the resulting scene to artists who can move individual meshes and re-tune individual lights instead of starting from a baked render. ## Quick Start Ask the agent to reconstruct the attached reference image into an editable Godot scene with separate object meshes and tunable lights using this reconstruction pipeline.

Frequently Asked Questions about lumera-single-image-scene-reconstruction

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I turn a single image into an editable 3D scene in Godot?

Use a structured parsing pipeline: a VLM extracts object 3D boxes and parametric light tuples from the image, per-object meshes are generated, and an assembly script writes a .tscn with separate MeshInstance3D and light nodes. A bounded refinement loop then verifies renders against the reference image.

What is the Lumera pipeline for image-to-3D scene reconstruction?

Lumera is a UE5 research pipeline (arXiv:2607.20889) that parses object boxes and parametric lights with VLMs, reconstructs per-object meshes, estimates an HDR environment, and refines results with a dual-agent loop. It has no public code, so this Skill documents the architecture and the Godot-adaptable parts.

Does this image-to-3D approach work without the original Lumera models?

The VLM box/light parsers, per-object mesh generator, and HDR estimator are external components you must supply, such as a hosted VLM prompted to emit the JSON tuples. For blockouts you can substitute BoxMesh primitives and a generated panorama sky, then swap in real meshes later.

When should I use a Gaussian splat instead of scene reconstruction?

Use a splat or depth-mesh bake when you only need a visual backdrop, since it is cheaper and looks as good. Avoid this pipeline for large unbounded outdoor scenes, where geometry drifts by roughly 17 m Chamfer-L2, or when exact light positions and intensities are critical.

Why are reconstructed light positions inaccurate?

Per-light localization is a known weak point: the paper reports F1 of 0.209 at 0.5 m and about 2.7x intensity error, though scene-level light count recall is 0.998. Expect the right number of lights in roughly right places, and plan for hand-tuning with log-scale intensity calibration.

How do I keep the reconstruction agent loop from running forever?

Set hard round limits for the geometry stage (T_g) and light stage (T_l), plus early termination when the verifier report triggers it. Use a sliding-window history of the most recent action/report pairs so 30+ rounds fit inside the VLM context budget.