Browser-native WebGPU inference, realtime simulation, and generated spatial systems in one inspectable workbench.
Kaminos is the browser-based authoring workbench where several independently useful systems meet. It keeps a spatial scene alive while local models run, receives their outputs as editable world matter, and turns runtime machinery that survives the workbench into reusable packages and focused model ports.
The repository spans four connected capabilities:
- Browser-native intelligence. A published WebGPU inference runtime and a growing family of spatial and generative model ports run as components of an interactive browser application.
- Live materials. Stateful fire, smoke, fluids, particles, and rendering processes remain visible and authorable while compute continues.
- Generated beings. Separate experiments connect authored morphology to generated form, reconstructed casts to mechanical control, and generated creatures to terrain-relative motion.
- A world kiln. Images, meshes, splats, motion, material fields, simulation state, and generated environments can be inspected, corrected, staged, and composed in one WebGPU workbench.
| Surface | What it demonstrates | Entry point |
|---|---|---|
| Kaminos workbench | Scene editing, generated assets and live materials in one authoring environment | Run locally or build an experiment |
| Browser combustion films | A stateful WebGPU fire material captured while being driven through ignition, contraction, chromatic change, extinction, and rebirth | Watch the captured studies |
| WebGPU inference kit | Reusable browser model applications, with shared-device lifecycle and cooperative scheduling | Try the models or build with the kit |
| Material Photograph | One photograph becomes a relightable surface, with synchronized original, depth, image-as-albedo and inferred-material views | Explore in your browser or use your photograph locally |
| Spatial Asset Kiln | The workbench architecture for generated assets, live routes, World Chambers, Preview Benches, and Smoke Offers | Read the architecture |
Local AI that leaves room for your application. Generate images, estimate materials, segment photographs and animate skeletons on the visitor's GPU. Use the kit independently of the Kaminos editor: model applications handle loading, progress and cleanup, while cooperative scheduling lets inference share the GPU with your renderer.
Try the browser models or build with Klein Image Studio. The kit guide has the full model catalog, npm setup and integration examples.
Develop visual features inside the shared workbench: contribute your objects, domain operations and controls while reusing the scene, camera, lighting and inspector. Human controls and agent calls operate the same experiment. Start with the runnable material study and contributor guide, then compose the capture and recording tools your investigation needs.
The architectural center of Kaminos is a simple product requirement: local AI should behave like part of an application, not replace the application with a modal wait.
In one measured M4 Max Chrome run, SHARP generated 1,179,648 Gaussian splats in 185.3 seconds while the full Kaminos fire volume continued to simulate on every frame in the same browser and on the same GPU. Across 21,818 foreground frame intervals, p95 and p99 were 9.3ms and 10.0ms; 40 intervals exceeded 33.3ms.
In a separate shared-device run, Stable Fast 3D produced a complete textured GLB in 41.9 seconds while servicing 3,644 host frames. Page frame intervals had a p99 of 9.7ms and a maximum of 92.4ms, and the output was byte-identical to the monolithic route's output.
Kaminos is the authoring surface. The
@kaminos/webgpu-inference-kit is the reusable
runtime extracted from the work required to keep that surface responsive while
substantial models execute.
These films were captured directly from the browser while one stateful WebGPU combustion material was being computed and authored, not prerendered.
The material carries its history through control changes. Existing momentum continues through contraction, acceleration, chromatic transition, changing source geometry, and renewed expansion. A broad burner can gather into a jet, retain the structure already in flight, and rebuild into another morphology without resetting the simulation.
Browser Combustion Films presents one complete composition, four authored transitions, and seven compact studies of color, structure, width, and state history.
The current multi-field simulation runs directly in the browser through WebGPU. Its boundary-fire renderer derives compact structural fields around the combustion front, then uses those fields to guide where the volumetric renderer spends work. The same material can run in the isolated screening room, inhabit arbitrary scene geometry, or remain active as the foreground workload around browser inference.
The fire began as an answer to the question that produced the inference kit: what should a local AI application do while expensive inference occupies the machine? It should stay alive.
One matched-image assay showed an authored low-frequency morphology edit influencing generated mass, stance, and support organization. Separately, one reconstructed cast was registered to a control armature and posed through ordinary controls after manual skinning. A different textured reconstruction consumed terrain-relative motion in a browser world.
These are three adjacent research lines, not one creature moving through a completed end-to-end pipeline. The continuing frontier is stronger editable control: returning distinctions authored before generation as durable controls that can survive later rounds of transformation.
The workbench receives images, meshes, splats, motion, material fields, simulation state, and generated environments as things that can be handled, not merely viewed. Current substrate includes:
- Three.js/WebGPU scene editing, cameras, lighting, and persistence;
- splat import, correction, crop, orientation, and sidecars;
- mesh/splat hybrid rendering and scene-context integration;
- motion generation, transposition, preview, and export experiments;
- browser-native fluid, particle, and volumetric material processes;
- saved world scenes and preview workbenches for generated environments;
- shareable asset previews and browser tools for inspecting generated output.
The architecture is documented in Spatial Asset Kiln. Splat correction and renderer-consumption contracts are documented in Splat Assets.
The longer product loop is already visible in the repository:
source -> model route -> live workbench -> generated matter -> authored world
| Path | Role |
|---|---|
index.html |
Main Three.js/WebGPU workbench, scene editor, rendering routes, and authoring controls |
serve.py |
Local server and asset-browse API for Kaminos and sibling model outputs |
webgpu-inference-kit/ |
Published runtime package, examples, integration guides, and contract tests |
docs/flame-atlas/ |
Public browser-combustion film collection and capture manifest |
models/ and pipelines/ |
In-tree experimental models and generated-asset pipelines |
docs/ |
Kiln, splat, structural-control, and route architecture |
tests/ |
Browser, runtime, scene, simulation, pipeline, and public-surface contracts |
artifacts/ |
Retained visual investigations and renderer studies |
Focused model ports live in their own repositories once their boundaries are strong enough to stand alone. Kaminos remains the place where those ports meet rendering, simulation, generated assets, and actual product behavior.
Serve the checkout:
python3 serve.py 8095Open the workbench in a WebGPU-capable Chromium browser:
http://127.0.0.1:8095/
The experiment guide starts from a shared-scene material comparison. To explore the current volume route:
http://127.0.0.1:8095/?kaminos_volume_smoke=1&volume_scene=tall_plume
The strongest current routes are developed and measured on Apple Silicon. Other WebGPU devices may expose different performance and feature boundaries.
Kaminos is an active research workbench. The inference runtime is published as
a reusable npm package; the combustion material, generated-world controls, and
spatial authoring surfaces continue to evolve in the workbench. Model ports live
either under models/ here or in linked repositories. The
model catalog
links to each current implementation and application guide; repository location
does not indicate maturity or integration support.
For technical or collaboration inquiries, contact Noah Lyons.
The Basin Maintenance Division monitors GitHub issues daily. Tell us what you tried and where you got stuck.