A CUDA-accelerated Nuke plugin for physically-based lens flare simulation. FlareSim traces rays through a real lens prescription to produce ghost reflections that respond correctly to source position, wavelength, and aperture shape — without look-up textures or artist-painted elements.
Tutorial: https://youtu.be/yEsBOQNG16Y
FlareSim is built on the foundational work of Eamonn Nugent (@space55 · 55.dev), whose original CPU-based lens flare renderer established the core ray-tracing approach, optical physics, and lens file format that this project builds upon. Without that work, FlareSim would not exist. The original codebase can be found at github.com/space55/blackhole-rt.
FlareSim takes any image as input and treats bright pixels as light sources. For each source it fires a grid of rays across the entrance pupil of the lens, tracing them through every optical surface. At two user-selected surfaces the ray reflects (instead of transmitting) — this is a ghost bounce pair. The reflected ray continues to the sensor plane, where its contribution is splatted onto the output buffer.
Because the tracing is done on the actual surface geometry, the resulting ghosts automatically exhibit:
- Correct position — ghosts land exactly where the optics dictate, including the characteristic axis-flipping behaviour when the source crosses the frame centre
- Chromatic aberration — each ghost is traced at multiple wavelengths; surfaces with AR coatings and dispersive glass bend R/G/B by different amounts
- Aperture shape — the ghost footprint inherits the polygonal or circular aperture set on the node
- Fresnel weighting — each bounce is weighted by the reflectance of its surface, including single-layer MgF2 AR coating simulation
The kernel runs entirely on GPU (NVIDIA CUDA). Render times are typically under one second per frame for standard HD/2K images.
- Copy
FlareSim.dllto a directory on yourNUKE_PATH, for example%USERPROFILE%/.nuke/plugins/. - Copy
menu.pyto the same directory. If you already have amenu.pythere, add this line to it instead:nuke.menu('Nodes').addCommand('Filter/FlareSim', 'nuke.createNode("FlareSim")')
- Copy the
lenses/folder somewhere accessible and point the Lens File knob at one of the.lensfiles. - Restart Nuke.
FlareSimwill appear in the node menu under Filter.
Requirements: Windows, Nuke 16, an NVIDIA GPU (any card from 2012 onward, compute capability 3.0+), CUDA runtime 13.
- Connect your beauty plate (or a sky/exterior shot with visible light sources) to FlareSim's input.
- Point Lens File at a
.lensprescription —doublegauss.lensis a good starting point. - Set FOV H to match the horizontal field of view of your camera.
- Increase Threshold until only the brightest highlights register as sources
(watch the
source.rgbchannels, or enable Show Sources). - Adjust Flare Gain to taste.
- Use the Pairs tab to disable any ghost pairs that are unwanted.
FlareSim writes several named channel groups in addition to passing the input RGB through:
| Channel | Contents |
|---|---|
flare.rgb |
Ghost reflections only |
source.rgb |
Map of detected bright sources (useful for threshold tuning) |
haze.rgb |
Veiling glare / scattered light glow |
starburst.rgb |
Diffraction spikes |
What appears in the main rgba channels depends on the Output Mode setting.
Flare Gain Overall intensity multiplier for the ghost reflections. Because ghost Fresnel weights are very small (two reflections, each typically 1–4%), a gain of several hundred is normal.
Ray Grid (NxN) Number of entrance-pupil samples per dimension. The kernel fires N² rays per (source, ghost pair) combination. Higher values produce smoother, less noisy ghosts at the cost of render time. 32–64 is a good starting range; 128 is high quality.
Pupil Jitter Sampling pattern for the entrance-pupil grid:
- Off — regular NxN grid. Fastest; can show a faint dot pattern at low ray counts.
- Stratified — one randomly-placed sample per grid cell. Breaks up the dot pattern. Use with Auto Seed for automatic per-frame variation.
- Halton — quasi-random low-discrepancy sequence. Deterministic (no seed needed), very even coverage. Good for still frames and motion-blur renders.
Jitter Seed Integer seed for the Stratified noise pattern. Only used when Auto Seed is off. Has no effect in Off or Halton modes.
Auto Seed When on (the default), the seed is automatically set to the current frame number. Sample positions change on every frame with no manual keyframing required — equivalent to animating Jitter Seed = frame. Turn off to use the fixed Jitter Seed value, which gives a repeatable pattern useful for still renders or multi-pass workflows where the noise must match exactly.
Threshold Minimum luminance (in scene-linear values) for a pixel to register as a light source. Raise this to suppress dim sources and reduce render time. Use Show Sources to see exactly what is being detected.
Source Cap
Clamps the luminance of any single source before it contributes to the flare.
Useful when one extremely bright highlight would otherwise dominate all ghosts.
0 = off.
Max Sources
Maximum number of bright sources passed to the GPU per frame. When there are
more sources than this limit, the N brightest (by luma) are kept and the rest
are discarded. 0 = unlimited.
Reducing this is the primary tool for managing render time on frames with many sources (e.g. sparkly water, city nights). It also prevents runaway render times if the threshold is accidentally set too low.
Downsample Stride used when scanning the input image for bright sources. A value of 4 samples every 4th pixel in each axis (1/16th of all pixels). Lower values find more sources and are more accurate but slower. For most shots 4–8 is sufficient.
Enable Preview Mode When on, the four settings below replace their counterparts from the Ghost and Spectral tabs for the render. This lets you dial in high-quality final settings once, then toggle a single button to drop into a fast interactive preview.
Ray Grid / Max Sources / Downsample / Spectral Samples (preview) Fast equivalents of the main settings. Suggested starting points: Ray Grid 16, Downsample 8, Max Sources 100, Spectral Samples 3 (R/G/B).
Brightness compensation
Preview mode automatically scales Flare Gain by (preview_downsample / final_downsample)² to compensate for the fact that a coarser downsample
detects fewer sources. Switching between preview and final should therefore
produce approximately the same brightness.
This compensation covers Downsample only. If the preview Max Sources limit is actively capping the source count (i.e. more sources exist than the limit), the preview will be slightly dimmer than the final. To avoid this, keep preview Max Sources equal to or higher than the final Max Sources value and rely on Downsample as the primary speed lever.
These settings tell FlareSim the field of view of the camera, which it uses to convert pixel positions into ray angles.
Use Sensor Size When enabled, FOV is computed from Sensor Width, Sensor Height, and Focal Length — matching the way most real-world cameras are specified. When disabled, enter the FOV angles directly.
Sensor Preset Fills Sensor Width and Sensor Height from a list of common formats (Super35, Full Frame, Micro Four Thirds, etc.).
FOV H / FOV V (deg) Horizontal and vertical field of view in degrees. Only active when Use Sensor Size is off.
Auto FOV V Derives vertical FOV from horizontal FOV and the image aspect ratio. Disable this only if your image has non-square pixels or a non-standard crop.
Sensor Width / Height (mm) and Focal Length (mm) Physical sensor dimensions and focal length. Only active when Use Sensor Size is on.
Aperture Blades
0 = circular aperture. 3–16 = regular polygon with that many sides.
The aperture shape directly affects the footprint of each ghost: a hexagonal
aperture produces hexagonal ghosts, matching what a real 6-bladed iris would
produce.
Aperture Rotation Rotates the polygonal aperture in degrees. Useful to match a specific iris orientation captured in footage.
Spectral Samples Number of wavelengths sampled per ray. More samples produce smoother colour fringing at the cost of render time.
| Setting | Wavelengths | Use |
|---|---|---|
| 3 (R/G/B) | 650 / 550 / 450 nm | Fast; one sample per channel |
| 5 | 400–700 nm, 5 steps | Good balance |
| 7–11 | Finer steps | High quality; smooth chromatic aberration |
Ghost Blur / Ghost Blur Passes
Applies a fast box blur to the ghost buffer after rendering. Radius is
expressed as a fraction of the image diagonal (e.g. 0.003 on a 1920-wide
image = ~7 pixels). Multiple passes approximate a Gaussian; 3 passes is
usually sufficient.
This is useful for adding a small amount of out-of-focus softness to the ghosts, or for hiding the point-splat nature of the renderer at low ray counts.
Haze Gain
Enables veiling glare — a broad, soft glow written to haze.rgb. Each source
paints its colour into a block of the haze buffer, which is then blurred with
a wide radius. This simulates light scattered inside the lens barrel.
0 = off.
Haze Radius / Haze Blur Passes Blur radius (as a fraction of the image diagonal) and number of passes for the haze glow. Larger radii produce a more diffuse, "milky" effect.
Starburst Gain
Enables diffraction spikes written to starburst.rgb. 0 = off.
The starburst is computed from the squared magnitude of the 2D FFT of the
aperture mask — the physically correct Fraunhofer diffraction pattern. Spike
count, angle, and falloff all emerge from the aperture geometry: a hexagonal
aperture produces 6 spikes, an octagonal aperture 8 spikes, and so on. A
circular aperture (0 blades) produces Airy rings rather than spikes.
The pattern is wavelength-dependent — red light diffracts more than blue, so the spikes have a blue centre and red tips. This colouration is computed per-channel (R/G/B at 650/550/450 nm) and requires no artistic tuning.
Typical gain values are 50–500.
Starburst Scale
Controls how far the spikes extend, expressed as a fraction of the image
diagonal. 0.15 means spikes reach roughly 15% of the diagonal from the
source — about 330 pixels on a 1920×1080 image.
Show Sources
Writes the detected source map into the source.rgb channels regardless of
the current Output Mode. Use this to tune Threshold and Downsample —
you should see the highlights you want to flare, and nothing else.
Output Mode
Controls which data appears in the main rgba channels:
- Separate channels — input RGB passes through unchanged; flare data is only
in
flare.rgb/source.rgb/haze.rgb. Use this when compositing manually downstream. - Flare as RGB — the ghost contribution is written into the main RGB as well
as
flare.rgb. The original input RGB is discarded. - Sources Only — shows the detected source map in the main RGB channels and skips all rendering (ghosts, haze, starburst) entirely. Use this to quickly tune Threshold, Downsample, and Max Sources without waiting for a full render — the node will respond almost instantly.
The Pairs tab lists every ghost bounce pair that the loaded lens produces, one
checkbox per pair. A pair is labelled (A, B) where A and B are the two
surface indices (0-based) at which the ray reflects.
FlareSim pre-filters the list and only shows pairs whose estimated Fresnel reflectance is above a minimum threshold, so very dim pairs are already hidden.
Refresh Pairs rebuilds the list from the current lens file and FOV settings. Use this if you change the lens or FOV and the pair list looks stale.
Select All / Deselect All toggle all pairs at once.
Disabling a pair completely removes it from the GPU workload, so unchecking pairs that don't contribute visually is a free render-time saving.
Lens files are plain-text .lens prescriptions. Each file describes one lens
as an ordered list of optical surfaces from front (entrance) to back (sensor).
name: Double Gauss 58mm f/2
focal_length: 58.0
surfaces:
# radius thickness ior abbe semi_ap coating
39.68 8.36 1.670 47.2 25.0 1
152.73 0.20 1.000 0.0 25.0 0
31.58 7.38 1.670 47.2 21.0 1
0 4.10 1.699 30.1 21.0 1
-1000.0 16.50 1.000 0.0 14.0 0
stop 12.13 1.000 0.0 12.0 0
...
| Column | Meaning |
|---|---|
radius |
Signed radius of curvature in mm. 0 = flat surface. |
thickness |
Axial distance to the next surface in mm. |
ior |
Refractive index of the medium after this surface (d-line, 587.6 nm). Use 1.0 for air. |
abbe |
Abbe V-number for dispersion. 0 = non-dispersive (air/vacuum). |
semi_ap |
Clear semi-diameter (half the clear aperture) in mm. |
coating |
AR coating layers. 0 = bare glass. 1 = single-layer MgF₂ quarter-wave. Higher values further reduce reflectance. |
The special keyword stop in the radius column marks the aperture stop surface.
Lines beginning with # are comments. The focal_length field is used to
convert FOV angles to physical sensor dimensions; set it to the nominal focal
length of the lens.
A large library of real lens prescriptions is included in lenses/lens_files/,
covering Nikon primes and zooms, Canon lenses, ARRI/Zeiss cinema lenses,
Cooke triplets, mirror lenses, and more.
FlareSim renders differently to standard Nuke nodes. A typical Nuke node processes one scanline at a time, so you see the image build progressively from top to bottom. FlareSim instead runs the entire render — source detection, GPU kernel, haze, starburst — on the first scanline request, stores the result in full-frame buffers, and then serves every subsequent scanline instantly from those buffers. The image therefore appears all at once rather than building progressively.
In practice this means there is a brief pause before anything is visible, then the full frame appears in one go. On fast renders (under a second) this is barely noticeable. On heavy frames — many sources, high ray grid, complex lens — the pause is longer. Use the render time printed to the console to track performance, and use Preview Mode for interactive work.
- Ray Grid is the biggest render-time lever. Start at 32 and raise to 64 or 128 only where needed.
- Threshold is the second biggest. Every source above threshold costs (n_pairs × n_grid²) ray traces. Fewer sources = much faster renders.
- Max Sources sets a hard ceiling on GPU work per frame, preventing surprise-slow frames on shots with many highlights.
- Disable pairs in the Pairs tab that don't contribute visually — they cost nothing when unchecked.
- Lenses with more elements produce more ghost pairs (an N-surface lens has N(N-1)/2 possible pairs). The doublegauss (11 surfaces) has 55; a complex zoom may have several hundred.
