Skip to content

Limitations

This document lists the current limitations with the most recent version of the Skewer Project.

Browser Support

The Scene Previewer uses the File System Access API (showDirectoryPicker, showOpenFilePicker) to read and write scene files directly from the local filesystem. This API is only supported in Chromium-based browsers (Chrome, Edge).

  • Firefox: Not supported at all.
  • Safari: Not supported at all.
  • Opera 72+: Supported. No known issues.

Scene Camera

The camera defined in scene.json supports keyframed look_from, look_at, vup, vfov, focus_distance, and aperture_radius values. The previewer can load, save, and scrub animated cameras, but it does not yet provide camera keyframe authoring controls.

The previewer viewport does provide interactive Three.js OrbitControls (orbit, pan, zoom) for inspection, but this does not affect the rendered output.

Scene Format

  • Geometry: Only three primitive types: sphere, quad, and obj mesh. No torus, cylinder, cone, NURBS, or other primitives.
  • Materials: Only three material models — lambertian, metal, dielectric. No subsurface scattering, cloth, hair, volume scattering, or custom BSDFs.
  • Media/Volumes: Only nanovdb type supported (.nvdb files). No procedural volumes, analytic fog, or heterogeneous media defined in-scene.
  • Mesh Import: Only Wavefront .obj format supported. No glTF, FBX, USD, or other interchange formats.
  • Integrators: Two choices — path_trace and normals. No photon mapping, bidirectional path tracing, VCM, or spectral MIS.

Rendering

  • CPU-only: Skewer is a CPU-based ray tracer. No GPU acceleration (though the data-oriented design was chosen to enable future GPU porting).
  • Spectral rendering: Uses 4 wavelength samples per ray. No RGB rendering path.
  • Deep segment limits: Maximum 16 deep segments per pixel, 16 depth buckets, 4 overlapping transmissive media.
  • Deep sample pool: Capped at ~64 chunks (~1.8 GB). Exceeding this silently drops samples with a warning.
  • No AOV system: No arbitrary output variables (albedo, normals, depth, etc. as separate channels). Each requires a separate render pass.
  • Spectral color: Wavelength sampling uses a temporary approximation, not tabulated spectral data.
  • Fireflies: Caustics, small emissive surfaces, and rough metals at grazing angles produce fireflies that cannot be fully eliminated. Reinhard tonemapping reduces but does not remove them.
  • No deferred differential geometry: Surface interaction differentials (dPdu, dPdv, etc.) are not implemented, which limits texture filtering quality.
  • Motion blur: Increases noise 2-4×, requiring significantly more samples.
  • Depth of Field: Increases noise noticeably.
  • Render time: Scales quadratically with resolution.

Animation

  • No looping or ping-pong: Animation clamps to the first and last keyframe values. No cycle modes.
  • Last keyframe curve ignored: The interpolation curve on the final keyframe has no effect (no next keyframe to interpolate toward).
  • Expanded BVH bounds: Animation requires expanded bounding boxes, which reduces acceleration structure efficiency.
  • Participating media is static: The engine does not support the animation of nvdb media. Adding this is currently a work in progress.

Cloud Platform

  • GCP-only: The distributed pipeline is native to Google Cloud Platform (Cloud Run, Cloud Workflows, Cloud Batch, GCS). No AWS, Azure, or on-premise backends.
  • 16-character layer ID limit: Layer IDs (derived from filenames) must be ≤ 16 characters. Longer names cause GCP Batch job creation to fail with INVALID_ARGUMENT.
  • No underscores in filenames: GCP Batch job IDs only allow lowercase letters, numbers, and hyphens. Underscores are converted but should be avoided.
  • SSD quota: Default 300 GB SSD limit caps concurrent VMs at ~10 (30 GB boot disk each).
  • Spot VM delays: Workers use SPOT instances by default. Provisioning may be delayed during high-demand periods.
  • Firebase authentication required: The previewer and API require Firebase Authentication with Google Identity Platform. No anonymous or local-only mode.
  • Rate limits: API limits /v1/jobs/init to 60/hr and /v1/jobs/{id}/submit to 5/hr per email by default.
  • File upload limits: Maximum 2048 entries per job init request; paths limited to 512 characters.
  • Data retention: Pipeline data retained for 30 days; cache for 90 days.

Build System

  • No pre-built binaries: Source build is required. No pre-compiled releases are available.
  • Toolchain requirements: C++20 compiler, CMake 3.21+, Go 1.21+, protoc, and several system libraries (OpenEXR, Imath, Zlib, libpng).
  • Platform-specific optimizations: Native CPU tunings (-march=znver3) only apply on Linux x86_64. Other platforms use a generic build.
  • Windows AVX2 requirement: Pre-built Windows binaries are compiled with /arch:AVX2 and require a CPU with AVX2 support (Intel Haswell 2013+ / AMD Excavator 2015+). If building from source on older hardware, pass -DSKEWER_BUILD_NATIVE_OPTIMIZATIONS=OFF to CMake. Systems without AVX2 must build from source.
  • Windows CRT conflict: MSVC CRT conflict workaround required (gtest_force_shared_crt) when building with tests.