Kaminos WebGPU Inference Kit

Generate images and run vision models locally in the browser. Start with a model application.

Try it in your browser

Generate an image with FLUX

Nine locally generated portraits of a blue toy train in different scenes and levels of jank — Generated in Chrome, 512 by 512 pixels with four steps.

Type a prompt and watch FLUX.2 [klein] 4B generate an image on your GPU. The text encoder, four diffusion steps and image decoder all run in the browser.

Generate an image Use it in your application

First visit: 3.2 GB of weights, cached for later runs. Use a WebGPU-capable browser.

Relight a photograph

A metallic glowing orb compared as a flat photograph, recovered depth, image-based relighting and inferred materials — One photograph, four synchronized views.

Explore three photographs as surfaces under changing light. MoGe provides the shape; SuperMat provides the surface finish. Compare them with the original photograph.

Explore depth & materials Use your photograph locally

Included model results are playable immediately. No model download.

Animate a skeleton with UniMate

Describe a motion, choose a skeleton or bring your own rig, and export the animation as a GLB. Text encoding and motion generation run on your GPU.

Generate motion Model guide

WebGPU required. Model weights download on first use; model terms apply.

Your hand, reconstructed live

Open, point and turn your hand. MobRecon reconstructs a 3D mesh from your camera, with hand detection and reconstruction running on your own GPU.

Try the hand demo Model setup

WebGPU and camera access required. One setup command prepares the models on your computer, then opens the demo. Model terms apply.

One runtime, different models

FLUX.2 [klein] 4B: Images from a prompt, with the full generation pipeline on WebGPU.; Model guide
TRELLIS 2: A textured 3D mesh from an image, with local tools for preparation and finishing.; Model guide
UniMate: Skeletal animation from a text prompt for different rigged skeletons.; Model guide
Hand reconstruction: Live hand meshes and landmarks from camera frames, using WiLoR and MobRecon.; Hand demo · Application guide
SuperMat: Albedo, roughness and metallic material maps from an image.; Explore materials · Port source
U2Net: Foreground cutouts from opaque RGB or RGBA images.; Application guide
SAM 3: Instance masks and transparent cutouts from an image and text prompt.; Example guide
MoGe: Depth, surface normals and point maps from a photograph.; Reveal depth · Model guide
Stable Fast 3D: A textured, UV-unwrapped GLB mesh from one image.; Model guide
SHARP: A Gaussian-splat scene from one image.; Model guide
Kimodo: Skeletal motion from a text prompt. Text embeddings use an external server.; Model guide

Build with the kit

Start with a model application for image generation, material maps, skeletal animation or background removal. It handles loading, queued inference, controls and cleanup. Share its GPU device with your renderer when you need one; model-specific options and lower-level access stay available.

npm install @kaminos/webgpu-inference-kit

Generate your first image Build with a model application Small rendering example Runtime and architecture API reference