Add pixel reconstruction filters via filter importance sampling - #122
Merged
Conversation
The renderer had no reconstruction filter at all: each sample jittered uniformly inside its own pixel footprint and contributed to that pixel alone — an implicit box filter of radius 0.5. This adds a selectable PixelFilter (box | triangle | gaussian | blackman | mitchell), chosen per scene with crust:pixelFilter / crust:pixelFilterRadius on the RenderSettings prim or per render with --filter / --filter-radius. Filtering is applied by filter importance sampling rather than splatting: each pixel warps its camera jitter through the filter's distribution (analytic inverse CDF for box/triangle, 64-bin tabulated |f| CDF for the rest) and weights the radiance by f/p; the film keeps a per-pixel weight sum and stores the weighted average. FIS is the right fit here because every mechanism in this renderer is strictly per-pixel — the adaptive early-stop, the OpenQMC per-pixel domain tree, and the inverse-variance pass blending would all break under a splatting film whose pixels share samples. Mitchell samples |f| and carries the sign in the weight, so its negative lobes stay unbiased; the CDF warp is monotone, so stratified QMC positions stay stratified. The default — box at radius 0.5 — evaluates to the identity jitter with weight exactly 1.0, so filterless scenes render bit-identically (pinned by box_half_radius_is_the_identity_jitter, and verified with check_images.sh against goldens from a clean pre-change build: all 12 sample scenes identical). Unit tests pin support/monotonicity, E[w] = 1 (an error here is an image-wide exposure drift), and that the weighted sample histogram reproduces each kernel; an integration test round-trips the USD attrs. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017iSyJtL8k3djCExtF2GrHV
The raster-to-NDC mapping divided by (width - 1) / (height - 1), which stretches the w-pixel grid across a viewport meant to be tiled by w pixels: a sub-pixel zoom of ~1/w that also let the last row and column sample past the top edge of the image plane (v > 1). Pixel i now covers [i/w, (i+1)/w) with its center at (i+0.5)/w, tiling [0, 1) exactly — the convention the camera's lower-left-corner + s·horizontal mapping expects. Every image shifts by up to half a pixel (cornellbox at 16 spp: 97% of pixels differ, mean abs diff 3.1e-3), so this is deliberately its own commit, separate from the bit-identical filter change — golden sets must be re-recorded across this boundary. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017iSyJtL8k3djCExtF2GrHV
Default reconstruction changes from box at radius 0.5 (the historical unfiltered jitter) to triangle at radius 1.0: proper antialiasing out of the box, at a well-understood sharpness cost and with the FIS weight still exactly 1. Scenes that author no crust:pixelFilter render differently from here on; box at radius 0.5 stays available (--filter box) and still reproduces the pre-filter jitter bit-identically, as the escape hatch for comparing against old renders — golden sets must be re-recorded across this commit. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017iSyJtL8k3djCExtF2GrHV
Contributor
Qodo reviews are paused for this user.Troubleshooting steps vary by plan Learn more → On a Teams plan? Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center? |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Implements pixel reconstruction filtering using filter importance sampling (FIS): instead of splatting samples into multiple pixels, each pixel warps its jitter through the filter's distribution and weights radiance by
f(x)/p(x). This keeps the renderer's per-pixel machinery (adaptive early-stop, per-pixel QMC domains, inverse-variance blending) intact while enabling smooth, configurable image reconstruction.Key Changes
New
filter.rsmodule: Defines five filter kernels (box, triangle, gaussian, blackman, mitchell) with configurable radii. Box at radius 0.5 is bit-identical to the historical unfiltered jitter, providing a comparison baseline. Mitchell's negative lobes are handled by sampling|f|and carrying the sign in the weight.FilterSamplerimplementation: Tabulates CDFs for filters without analytic inverse CDFs (gaussian, blackman, mitchell); box and triangle use closed-form solutions. The sampler maps uniform[0,1)variates to(offset, weight)pairs, maintaining monotonicity so stratified/QMC sequences stay stratified after warping.Integration into
Renderer: Modifiedrender_pixel()to apply filter importance sampling to camera jitter. The pixel estimate becomes a weighted averageΣwᵢ·Lᵢ / Σwᵢ, which reduces to the plain mean for box and triangle (where all weights are 1.0).USD scene support: Added
crust:pixelFilter(token) andcrust:pixelFilterRadius(float) attributes to RenderSettings, parsed inusd_import.rs. Defaults to triangle at radius 1.0.CLI flags:
--filter(box|triangle|gaussian|blackman|mitchell) and--filter-radius(pixels) override scene settings. The filter radius is measured from the pixel center; each filter has conventional defaults (box 0.5, triangle 1, gaussian/blackman 1.5, mitchell 2).Comprehensive tests: Verify bit-identical behavior for box at 0.5, monotonicity and support bounds, weight normalization (E[w] = 1), and that weighted histograms reproduce the kernel itself.
Implementation Details
f(x) · Σ|f| / (|f(x)| · Σf), ensuring unbiased reconstruction while handling negative contributions correctly.(pixel + jitter) / (resolution - 1)to(pixel + jitter) / resolution, which correctly tiles[0, 1)and prevents sampling past the image boundary.Σ(w·L) / Σwto handle variable weights; degenerate cases (e.g., all samples on Mitchell's negative lobes) fall back to the plain mean.The filtering is applied at zero cost per sample and preserves adaptive sampling's per-pixel operation.
https://claude.ai/code/session_017iSyJtL8k3djCExtF2GrHV