메뉴
Image quality analysis, artifact detection, convergence diagnostics, noise decomposition, and parameter corrections. Use when: diagnosing image quality problems, running ppt_diagnose, analyzing convergence, identifying bottlenecks, tuning render parameters via feedback, or comparing A/B renders.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
C++ code quality standards: style, naming, comments, structure, brevity. Use when: writing new C/C++/CUDA code, reviewing style, refactoring for readability, cleaning up comments, or enforcing project conventions. Applies to all .cpp, .h, .cu, .cuh files.
Multi-stage pipeline orchestration using 3 executables (ppt_analyze → photon_tracer → ppt_diagnose). Use when: wiring the analyze→render→diagnose pipeline, setting up preview or production rendering, coordinating skill invocation order, configuring render parameters, or debugging stage-to-stage data flow.
GPU post-processing pipeline inside photon_tracer: firefly filter (median+MAD), bloom (mip-chain Gaussian), tonemap (ACES+sRGB gamma), OptiX denoiser. Use when: fixing firefly artifacts, tuning bloom thresholds, adjusting tonemapping, debugging output corruption, integrating OptiX denoiser, or optimizing post-FX performance.
GPU render pipeline internals. Use when: modifying renderer source code, fixing material/intersection/energy bugs, or debugging noise/fireflies in implementation. NOT for simply running the executable (use orchestrator skill).
Scene loading, PBRT/OBJ parsing, material mapping, scene classification, and auto-tuning parameter recommendation. Use when: loading scenes, fixing parsing bugs, adjusting scene classification, tuning emitter detection, modifying SceneProfile, or running ppt_analyze for pre-render analysis.
Interactive scene editing and manipulation. Use when: modifying scenes in the editor, adding/removing objects, adjusting materials interactively, or debugging scene construction.
| name | quality |
| description | Image quality analysis, artifact detection, convergence diagnostics, noise decomposition, and parameter corrections. Use when: diagnosing image quality problems, running ppt_diagnose, analyzing convergence, identifying bottlenecks, tuning render parameters via feedback, or comparing A/B renders. |
The ppt_diagnose executable and in-render diagnostics system.
| Location | Files | Purpose |
|---|---|---|
src/diagnose/ | diagnose_main.cpp | ppt_diagnose entry point: loads EXR, runs analysis, prints JSON |
src/diagnose/ | image_oracle.h/.cpp | Main analysis engine: no-reference and reference-based |
src/diagnose/ | artifact_detector.h | 6 artifact types with severity scoring |
src/diagnose/ | noise_analyzer.h | Laplacian variance, spatial autocorrelation |
src/diagnose/ | image_metrics.h | PSNR, SSIM, RMSE computation |
src/diagnose/ | image_verdict.h | Verdict + ArtifactReport + Correction structs |
src/diagnose/ | diagnostics.h/.cpp | In-render convergence analysis |
src/diagnose/ | variance_tracker.h/.cpp | Per-pixel variance tracking from GPU buffers |
src/diagnose/ | convergence_analyzer.h | Convergence rate estimation (log-log fit) |
src/diagnose/ | bottleneck_report.h | Bottleneck identification + parameter recommendations |
ppt_diagnose <render.exr> [render_log.json]
Loads rendered EXR, runs image oracle, prints JSON to stdout:
{
"quality": { "noise_level", "convergence_rate", "summary" },
"artifacts": [{ "type", "severity", "description" }],
"corrections": [{ "target", "parameter", "current", "recommended", "rationale" }]
}
| Type | Detection | Threshold | Correction Target |
|---|---|---|---|
| Firefly | Pixel > K × local median | K=3–6 (adaptive per SPP) | path-integrator (clamp), post-processing (filter) |
| Splotch | Spatial autocorrelation ρ(lag=8) > 0.15 | 0.15 | path-guide (more training iters, lower split threshold) |
| HighNoise | Laplacian variance (Immerkaer 1996) | σ > scene-dependent | guide_fraction increase, more SPP |
| EnergyLoss | Mean luminance < expected (over-clamping) | 15% deviation | raise clamping thresholds |
| Banding | Gradient quantization | histogram peaks | increase precision / dithering |
| SlowConvergence | Rate α < 0.3 in log-log fit MSE∝N^(-α) | α < 0.3 | guide, photon, or scene-specific tuning |
photon_tracer --sweep renders at multiple SPP levels (default: 16,64,256,512,1024,2048), saves checkpoint images and a _sweep.json report with per-level noise + mean luminance. Use to diagnose noise reduction effectiveness across sample counts. See orchestrator skill for full details.When diagnosing artifacts or evaluating parameter changes, vary one parameter at a time. Render multiple values of that single parameter (ideally as --sweep runs), compare metrics, then lock the best value before moving to the next parameter. Changing multiple parameters simultaneously prevents attributing improvements or regressions to their cause. Only combine parameters after their individual effects are understood. See the orchestrator skill's "Single-Variable Isolation Strategy" for the full workflow.
ppt_analyze scene.pbrt > profile.json
photon_tracer --config profile.json scene.pbrt → output.exr + render_log.json
ppt_diagnose output.exr render_log.json > report.json
# Apply corrections from report.json → re-render (up to 3 iterations)