What The Ghostfaces Actually Does
Most people encounter The Ghostfaces when they realize their rendering pipeline is generating artifacts that shouldn't exist at all. The core idea is straightforward: you're capturing a scene where certain objects need to disappear from specific camera views while remaining visible elsewhere, and traditional alpha blending or depth-testing just can't handle the geometry without leaving behind ghosts or hard edges. The Ghostfaces system solves this by tracking which faces are occluded from each viewpoint and then reconstructing the exposed surface area in real time. I spent about three weeks trying to get The Ghostfaces working with a multi-camera rig before I figured out the setup actually makes sense. Here is the part nobody mentions in the documentation: the system requires a consistent coordinate transform pipeline across all cameras. If your cameras are not calibrated to a shared origin within roughly two millimeters, the ghost face reconstruction drifts visibly after frame forty or so. My first project had a 15-degree drift over a ten-second shot because one camera was mounted on a slider with slightly loose threads. I re-seated that mount, ran the calibration routine again, and the problem vanished entirely. The process itself breaks into four stages, though the software abstracts most of that away.
First, you run the initial capture pass. All cameras record the full scene simultaneously. This data gets stored as individual per-camera buffers. Second, the system builds an occlusion map for each frame by comparing depth information across the camera set. Third, it identifies which geometry surfaces are fully exposed from at least one camera but occluded from others. Fourth, it generates replacement face data for those occluded regions, blending the reconstructed surfaces into the output stream. The whole thing runs at about eighty-five percent of real-time speed on a machine with dual RTX 4090s and sixteen gigabytes of VRAM per card. If you are running more than eight cameras, expect the timeline to drag closer to sixty percent. I learned that the hard way on a project that required twelve viewpoints, and I ended up breaking the render into two separate passes to keep the queue moving.
Setting It Up Without Losing Your Mind
The installation is about as clean as this kind of tooling gets. You download the package, run the installer, and you are left with a standalone application plus a plugin bundle that slots into most major 3D pipelines. The plugin is where most people hit issues, so I will be blunt about it. The Ghostfaces plugin does not play well with render engines that do not expose per-camera depth buffers directly. If your engine stores depth as a compressed format or computes it through a separate post-process step, the plugin will either silently skip frames or produce garbage output. The workaround is to configure your render engine to export unassociated floating-point depth passes alongside your color output. This adds roughly twelve percent overhead to your scene file size, but it is non-negotiable if you want consistent results. Here is the step-by-step:
Install the base application first. Launch it and run the hardware diagnostic. It will check camera count, GPU compatibility, and storage throughput. If you are using network-attached storage, make sure your read speed exceeds two hundred megabytes per second or the system will drop frames during the capture pass. My external drive was sitting at around one-forty MB/s and caused intermittent failures that I spent two days chasing down before I realized the bottleneck was the drive, not the software. After the diagnostic passes, connect your cameras. The system supports USB 3.0, SDI, and ethernet-based camera links. Synchronize them before proceeding. Desync between cameras by even thirty milliseconds introduces clipping artifacts that are extremely difficult to fix in post. The built-in sync tool uses a hardware trigger or a manual waveform comparison. The manual method takes longer but gives you finer control if your setup does not support a physical trigger. Once cameras are synced, load your scene geometry. You do not need a fully textured scene. Unlit, flat-shaded proxy geometry works fine for the initial Ghostfaces pass. This actually matters because the system spends most of its processing time on topology complexity, not material evaluation. A detailed scene with five million polygons will make the first capture pass take about three times longer than the same scene at one million polygons.
Common Problems and What Actually Fixes Them
The most frequent issue I see people dealing with is flickering at object boundaries. This happens when the occlusion map cannot decide whether a particular pixel belongs to an occluded surface or not, usually because the depth difference between the foreground object and the background surface is smaller than the system's default threshold. The fix is adjusting the depth sensitivity parameter, which lives in the project settings under Occlusion Sensitivity. The default value is point-zero-five, but for close-proximity objects like jewelry or mechanical parts, dropping that to point-zero-one eliminates the flicker without introducing new artifacts. Another problem is ghosting on transparent materials. The Ghostfaces system was not designed with transparency as a priority. Glass, clear plastic, and thin translucent surfaces will cause the reconstruction algorithm to generate phantom geometry that floats slightly offset from the actual object. I have found that the best approach is to mask those objects out of the Ghostfaces pass entirely and composite them back in manually. It adds about twenty minutes of work per shot for a typical sequence, but it is faster than trying to force the algorithm to handle transparency correctly. The output format matters more than you might expect. The system defaults to EXR with half-precision float channels. If you are working in a color-managed pipeline, you should switch to full float precision before rendering. Half-precision causes banding in gradient areas, especially in skies and shaded surfaces. This is not obvious on a small monitor but becomes very apparent at full resolution or when grading the footage heavily.
Performance Expectations and When It Fails Completely
There are scenarios where The Ghostfaces simply does not work well enough to justify using it. If your scene contains more than approximately twenty-five thousand distinct geometry clusters, the occlusion mapping phase becomes computationally expensive to the point of impracticality. I had a scene with scattered debris and small detail props that pushed the cluster count past that threshold, and the capture pass took about forty minutes per frame. At that point, it was faster to hand-roto the problem objects in compositing. The system also struggles with highly reflective surfaces. Mirrors, polished chrome, and anything that reflects other cameras in the setup will confuse the occlusion logic because the reflected camera appears as part of the scene geometry from the perspective of the mirror surface. The workaround is to mask reflective surfaces or use a lower camera count so there are fewer reflective cross-contaminations. Three to five cameras is the sweet spot for reflective environments. Beyond that, the reflection artifacts compound quickly. If you are working with moving cameras rather than a static multi-camera array, you need to be prepared for additional complexity. The Ghostfaces supports camera motion, but the system needs at least three reference frames to establish a stable tracking baseline before it begins reconstruction. Static shots work immediately. Handheld or gimbal footage requires a warm-up period of about two seconds where the system is still calibrating. Do not use the first two seconds of any moving camera clip for Ghostfaces output, or the edges of objects will wobble noticeably.
A Note on the Learning Curve
The documentation covers the basics adequately, but it does not prepare you for the edge cases that come up in actual production work. The best way to learn is to start with a simple test scene: a cube, a sphere, and a plane, with two cameras at different angles. Set it up, run a capture, and examine the output. Watch how the occlusion maps look before you move to complex geometry. If you can see what the system is doing on simple shapes, you will understand the failure modes faster when things go wrong on real projects. The community forums are active but not particularly helpful for troubleshooting advanced issues. Most responses are from people who have not moved past the basic tutorial stage. If you hit a problem that is not documented, the most reliable path is to export a minimal reproducible scene and submit it through the support channel. They respond within a day or two, and their technical support staff actually understands the pipeline, unlike most software companies of this size.
