Render Fidelity Comparison Results

The purpose of glTF is to standardize Physically-Based Rendering (PBR) materials such that you can be confident your model will appear as intended in any lighting environment in any renderer. This is a very ambitious goal, as real-time rendering at this level of quality is still very much an area of active research with improvements being made constantly. This site demonstrates where we are on that path to convergence and highlights areas that could still use improvement. We are comparing the most popular web renderers: three.js (represented by <model-viewer>), filament.js, babylon.js, gltf-sample-viewer, and three-gpu-pathtracer. If any other renderers would like to be included, please open a PR adding them to the render-fidelity-tools package.

In addition, we are showing a ground-truth render provided by Dassault STELLAR, which is a professional path-tracing renderer and approaches the physical result with far fewer approximations than are necessary in real-time renderers.

The first set of scenarios are for demonstrating material capabilities, so these are primarily fidelity tests. Following these are simpler cases that test glTF spec conformance, but are not particularly interesting for fidelity. At the bottom are the Furnace Test and the Directional Light Test. The Furnace Test checks energy conservation and the correct result is the whole 7x7 grid of metal-rough spheres being invisible (indistinguishable from the background). The top (metal) and bottom (dielectric) rows are passing, while the mixtures lose some energy. The Directional Light Test checks the worst-case scenario for environmental lighting approximations: the case where a single bright pixel represents a directional light. The Filament version uses an actual directional light, so this is the ground truth.

In <model-viewer>, we do not consider rendering changes to be breaking changes, as our quality is incrementally improving with every release. However, the difference between the renders you see here bounds how much you can expect our rendering to change going forward. Note the largest differences are in the handling of transparent materials, as this is quite difficult to get right, so the largest changes will likely be here.

We show a simple logarithmic metric on an average of the pixel differences, excluding the transparent background. This is not a perceptual metric, so take it with a grain of salt. It is mostly there to help us identify regressions.