How do I install the Spark Scene skill?

Run npx skillmds@latest add dreamlab-ai/spark-scene in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Spark Scene safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Spark Scene?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Spark Scene free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Spark Scene?

dreamlab-ai (@dreamlab-ai) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Spark Scene by dreamlab-ai — Agent Skill

Name: Spark Scene
Author: dreamlab-ai

Spark Scene

Integrate Spark Gaussian splats into a Three.js browser scene with spatial annotations, evidence links, or graph overlays. Use for viewing trained captures alongside data, splat picking, and scene registration; use lichtfeld-studio for reconstruction/training and blender for mesh authoring.

dreamlab-ai b558396 2 files · 7.8 KB Updated Sep 17, 2026

File contents

Spark scene integration

Use Spark as the browser presentation layer for an existing trained scene. It complements the LichtFeld/COLMAP capture pipeline; it does not recover geometry, metric scale, semantics, or reliable evidence from images by itself.

Read the integration reference when adding a scene layer. It covers package pins, the shared render loop, coordinate registration, annotation ownership, cleanup, and the hardware acceptance checks.

Fit and boundaries

Add splats to the application's existing Three.js scene and renderer. Preserve its camera, selection, graph overlays and XR lifecycle. Avoid a second canvas or animation loop merely to display the capture.
Treat spatial coordinates and semantic embedding layouts as distinct frames. An embedding neighbour is not a nearby physical object. Place data into the capture only through an explicit, recorded spatial anchor or registration.
Keep evidence/annotation records outside the splat buffer. Splat indices can change during conversion, level-of-detail selection and retraining; use stable application identifiers to link records and geometry.
Reuse the host's identifier allocator and durable-state APIs. In Agentbox, durable identifiers come from management-api/lib/uris.js, and persistence uses the configured adapters. The viewer does not introduce a memory store.
For reconstruction, export a trained PLY/SPZ from lichtfeld-studio; for mesh modelling use blender. Use browser for real WebGL2 validation. Loading this skill requires neither the CUDA training gate nor a new MCP daemon.

For an investigative scene, display source capture references, measured scale, registration error and reconstruction gaps. Label inferred placements separately from observed ones. A photorealistic rendering does not establish measurement accuracy or resolve occluded surfaces.

---name: spark-scenedescription: Integrate Spark Gaussian splats into a Three.js browser scene with spatial annotations, evidence links, or graph overlays. Use for viewing trained captures alongside data, splat picking, and scene registration; use lichtfeld-studio for reconstruction/training and blender for mesh authoring.---# Spark scene integrationUse Spark as the browser presentation layer for an existing trained scene. Itcomplements the LichtFeld/COLMAP capture pipeline; it does not recover geometry,metric scale, semantics, or reliable evidence from images by itself.Read [the integration reference](references/integration.md) when adding a scenelayer. It covers package pins, the shared render loop, coordinate registration,annotation ownership, cleanup, and the hardware acceptance checks.## Fit and boundaries- Add splats to the application's existing Three.js scene and renderer. Preserve its camera, selection, graph overlays and XR lifecycle. Avoid a second canvas or animation loop merely to display the capture.- Treat spatial coordinates and semantic embedding layouts as distinct frames. An embedding neighbour is not a nearby physical object. Place data into the capture only through an explicit, recorded spatial anchor or registration.- Keep evidence/annotation records outside the splat buffer. Splat indices can change during conversion, level-of-detail selection and retraining; use stable application identifiers to link records and geometry.- Reuse the host's identifier allocator and durable-state APIs. In Agentbox, durable identifiers come from `management-api/lib/uris.js`, and persistence uses the configured adapters. The viewer does not introduce a memory store.- For reconstruction, export a trained PLY/SPZ from `lichtfeld-studio`; for mesh modelling use `blender`. Use `browser` for real WebGL2 validation. Loading this skill requires neither the CUDA training gate nor a new MCP daemon.For an investigative scene, display source capture references, measured scale,registration error and reconstruction gaps. Label inferred placements separatelyfrom observed ones. A photorealistic rendering does not establish measurementaccuracy or resolve occluded surfaces.

Frequently asked questions