three-ntc

Neural Texture Compression runtime for three.js — load .ntc files and get back a standard MeshPhysicalNodeMaterial (via TSL) that decodes a compact grid + MLP model on the GPU.

An implementation of NVIDIA's 2023 Neural Texture Compression paper: NTC jointly fits a material's whole texture stack — albedo, normal, roughness, metalness, and more — into a single grid + MLP model instead of one texture per channel. Every material pictured above is about 93KB total, often smaller than a single texture from the original material, for a fraction of the memory and download cost of standard PBR textures.

Try it live at three-ntc.ben3d.ca.

.ntc models are trained with the companion three-ntc-trainer package.

Install

npm install three-ntc

Usage

import { NTCLoader, NTCNodeMaterial } from 'three-ntc';

const loader = new NTCLoader();
const { cpuModel, channelClassification } = await loader.loadAsync('gold.ntc');

const material = new NTCNodeMaterial(cpuModel, channelClassification);
const mesh = new THREE.Mesh(geometry, material);

NTCNodeMaterial extends three.js's MeshPhysicalNodeMaterial, so it drops straight into an existing WebGPU renderer scene alongside ordinary materials.

Runtime sampling

const material = new NTCNodeMaterial(cpuModel, channelClassification, {
  samplingMode: 'nearest', // default
  lodBias: 0,
});

// Sampling changes rebuild the TSL graph and compile a specialized shader on next use.
material.setSamplingMode('stochastic'); // or 'nearest' / 'trilinear'
material.samplingMode = 'nearest';     // equivalent property setter
material.setLodBias(1);                // live uniform; positive bias selects finer mips

| Mode | MLP evaluations per material sample | Behavior | | --- | --- | --- | | nearest | 1 | Reconstruct the nearest physical texel at the nearest mip. | | stochastic | 1 | Randomly select a texel and mip with trilinear sampling probabilities. Noise varies by screen pixel and frame. | | trilinear | 8 | Reconstruct and blend four texels at each of two mip levels. |

Sampling mode is a JavaScript build-time choice, not a shader uniform. Each shader contains only its selected sampling path. Changing the property or calling the setter rebuilds the material's channel nodes and marks it for recompilation. Selecting the same mode again does nothing. Existing grid textures, decoder parameters, UV/LOD settings, and debug view are retained; training updates still upload in place.

Stochastic sampling follows the paper's UV/LOD jitter approach. It introduces noise; this library and the website currently apply no temporal reconstruction. Its expected material-channel values match trilinear filtering, but averaging shaded samples is not equivalent to shading averaged material channels. Add a temporal resolve in the renderer when using this mode for stable images.

The old interpolation option and setInterpolation() have been replaced by samplingMode and setSamplingMode(). Sampling controls runtime reconstruction, not the training model or its learned G0/G1 feature interpolation. Settings are not stored in .ntc exports. setLodBias() affects automatic LOD only; an explicit constructor lodNode takes precedence. The website viewer and trainer preview both expose sampling and LOD bias controls and start with nearest sampling.