Interactive Mode¶

Modes 3 (CUDA) and 5 (OptiX) open an SDL2 window where any scene can be explored in real time: mouse-driven camera, progressive accumulation, and a Dear ImGui panel holding every rendering parameter. Mode 5 is the default when the OptiX SDK is available; it is also the one with the motion regulator and the AI denoiser.

Starting interactive mode¶

cd build
./rayon --scene ../resources/scenes/17_cloud_bunny.yaml        # mode 5 (or 3 without OptiX)
./rayon -m 3 --scene ../resources/scenes/09_color_bleed_box.yaml

Camera controls¶

Any camera change restarts the accumulation.

Keyboard shortcuts¶

How accumulation works¶

Each frame adds a batch of Samples/Batch samples to every pixel (8 by default; a scene can set settings.samples_per_batch, and --samples-per-batch on the command line wins). The samples go into a float accumulation buffer that stays on the GPU; a small kernel converts it to the 8-bit display image.

While the camera is still, the renderer keeps adding batches until one of these happens:

• Adaptive sampling reports that 99.9 % of the pixels have converged (see below),
• the budget of 100 000 samples per pixel is reached,
• you pause it with Space, or started with --no-auto-accumulate.

While the camera moves, the OptiX renderer trades quality for frame rate (see Camera Motion below). The CUDA renderer keeps rendering full batches.

The control panel¶

The panel sits at the right of the window. Every section is a collapsible header.

Scene¶

The scene combo lists every YAML file found in resources/scenes/; the arrow keys cycle through it. The current camera position and look-at point are displayed, and the button Save camera, lights & tone map to scene YAML writes them back into the file the scene was loaded from, together with the light settings and the tone map. Only the camera: lines (position, look_at, fov) and the settings: keys background_intensity, light_intensity, light_falloff, tone_mapping, exposure (and indirect_clamp on OptiX) are rewritten; comments and everything else are kept. This is how the bundled scenes got their viewpoints.

Performance Monitoring¶

Samples per pixel so far, throughput in samples per second, time per pass, and a history graph of both. Two diagnostic maps replace the image while checked (one at a time):

Camera Settings¶

Field of view, depth of field (aperture and focus distance), auto orbit.

Environment & Materials¶

The light and sky values are live: they are launch parameters, no scene rebuild is needed.

Procedural Patterns¶

Shown when the scene holds a golf ball (displaced sphere): number, radius and depth of the dimples.

Cloud / Smoke¶

Shown when the scene holds a volume material: albedo, density, phase asymmetry, noise scale and strength, edge, lumps, billow. Every change re-bakes the density field in a few milliseconds (Volumes).

Adaptive Sampling¶

Pixels lit only by rare bright paths are protected from freezing too early. Disable with --no-adaptive-sampling. The map of the samples spent per pixel is the Sample Heatmap of the Performance Monitoring section.

MIS / NEE¶

Enable MIS (master toggle for next event estimation and the MIS weights), NEE stride (beyond the first bounce, shadow rays on 1 of every N samples, 1 by default) and the Sobol sampler checkbox (uncheck for PCG). See Multiple Importance Sampling.

Camera Motion (OptiX)¶

Everything in this section applies only while the camera moves. The accumulation restarts on every move, so none of it changes the final image.

The readout under the sliders shows the resolution fraction and the samples used by the last motion frame, and the denoiser time when it runs.

Appearance¶

Theme (light, classic, nord, dracula, gruvbox, catppuccin), logo, samples-per-second counter.

Controls & Help¶

The mouse and keyboard reference above.

Display pipeline¶

flowchart TD
    A["Path tracing kernel\n(CUDA or OptiX)"] -->|adds a batch of samples| B["float4 accumulation buffer\n(device memory)"]
    B -->|gamma kernel, upscale of motion frames| C["uint8 display buffer\n(device memory)"]
    C -->|async copy to pinned host memory| D["uint8 host buffer"]
    D -->|SDL_UpdateTexture| E["SDL texture"]
    E -->|SDL_RenderCopy| F["Window"]

Only three bytes per pixel cross from the device to the host each frame, about 6 MB at 1920×1080. The float accumulation buffer never leaves the GPU.

Sample screenshot¶

The interactive window after a few seconds of stillness. The panel on the right shows the sample count, the throughput graph and the live sliders.