A model does not learn the physical world from a single image, a single angle, a single light source, or a single condition. It learns through difference. Through repetition. Through motion. Through uncertainty. Through the controlled exposure of form under changing circumstances.
Each generated image becomes a controlled variation. Each variation gives a model another way to learn
For machine perception, these variations are not noise. They are the field
The world rarely presents objects under perfect conditions, rain, shadow, glare, motion, distance, reflection, occlusion; seen from above, below, sideways, partially hidden, or in motion.
Spectrum exists to make variation available. It gives builders a way to generate visual data across controlled differences in texture, lighting, angle, movement, and environment.
The purpose is not to replace the real world. The purpose is to help models encounter more of its possible appearances
A Unity-based synthetic data generator for computer-vision datasets
It allows an image to be applied to a specific part of a 3D object rather than the entire object, the rest of the object remains visually intact
Segmented application
A texture may be placed on a door. A hood. A panel. A surface. A tool. A material region. A segmented portion of a larger form. This gives the user more control over how variation appears in the scene, specific surfaces change while the whole object remains coherent.
3D environments
3D environments, segmented objects, and user-uploaded image textures.
Randomization
Lighting randomization, camera movement, material variation, and object motion.
Structure preserved
The result is a dataset that can preserve structure while varying appearance, diverse image datasets for machine learning pipelines.
Not as a static image. As a changing visual condition
Spectrum generates synthetic image datasets with controlled diversity. Each dataset can include variations across:
Real-world objects are rarely visually uniform, drag to compare
Spectrum applies the image to a designated region of the object. The image becomes part of the object's visual field without overwhelming its structure. This supports more realistic and more controlled dataset generation.
Variation is the discipline of Spectrum
A dataset becomes useful when it teaches a model to recognize through change. Spectrum introduces controlled randomization across the visual scene
Surface & material
Spectrum can vary surface properties such as roughness, glossiness, and reflectivity. The same texture may appear different when placed on a matte surface, a reflective surface, a worn surface, or a polished surface. This helps models learn beyond a single material condition.
Lighting
Spectrum can vary lighting intensity, direction, and color. This allows datasets to include soft light, harsh light, shadow, glare, low light, and color-shifted environments. For perception systems, lighting is not cosmetic. Lighting changes what can be seen.
Camera
Spectrum can vary camera position, distance, and angle. A model trained only from one viewpoint becomes fragile. A model trained across many viewpoints begins to understand form. Camera randomization helps expose the object from different perspectives, allowing the model to learn through orientation rather than memorization.
Motion
Spectrum can vary object movement and rotation. Objects do not only exist. They move. They turn. They enter and leave the frame. They appear at unstable angles. They are seen in transition. Motion randomization helps create datasets that reflect the instability of real environments.
A single image can become many appearances
A single object can be rendered across many conditions. A rare case can be simulated before it is encountered again. This is how a scarce sample becomes a usable distribution.
In many real scenarios, there are not thousands of examples available
Traditional machine learning systems often struggle under these conditions because they do not have enough variation to learn from. Spectrum expands a limited starting point into a broader field of visual possibilities. This makes Spectrum useful for research environments where data is scarce but model performance still matters.
The deeper goal is not simply to create more images. It is to create more conditions of seeing
Machine perception fails when it has only learned the easy version of reality, clean lighting, centred objects, familiar angles. Then the world changes.
- A shadow appears.
- A surface reflects.
- A camera moves.
- A texture shifts.
- An object rotates.
- A background becomes complex.
- The distance changes.
- The scene becomes unstable.
Spectrum exists for this instability. It gives researchers a way to create visual variation before failure occurs in the field.
Read: Measurement Carries Responsibility →A model becomes more resilient when it has learned across conditions
A generated image should not only exist. It should carry a record of how it was created
Spectrum generates datasets at any scale, exported as PNG or JPEG with metadata describing the conditions under which each image was made, because synthetic data is most useful when it is traceable.
Metadata allows researchers to reproduce, compare, audit, and refine the generation process.
It makes the conditions of perception visible
Spectrum belongs inside Voir because synthetic visual generation is not separate from spatial intelligence. It is one of the ways spatial intelligence is studied
- The same object is not the same image from every angle.
- The same surface is not the same under every light.
- The same form is not the same in motion.
- The same scene is not the same from every position.
To study perception, we must study variation. To study variation, we need instruments that can generate it. Spectrum is one such instrument. It helps expose the relationship between object, surface, camera, light, motion, and environment.
Read: From Symbolic to Spatial Awareness →An instrument, not a completed theory of visual reality
Controllable, and generated
Synthetic images are useful because they are controllable. They are limited because they are generated. Some scenes may require manual adjustment. Some object segmentations may require refinement. Some lighting or motion conditions may appear unnatural. Some real-world complexity cannot yet be fully simulated.
These limits are not hidden
These limits matter. They are not hidden. They are part of the research.
Honest about the distance
The purpose of Spectrum is to create useful controlled variation, while remaining honest about the distance between simulation and reality. That distance is itself one of the central questions of spatial intelligence.
Not merely a generator of images. A generator of conditions
Spectrum is Voir's first public instrument. Prysm gives the research framework. Spectrum gives the field a generator. It turns questions of perception into controlled visual conditions, and allows researchers to ask:
- What changes when light changes?
- What changes when angle changes?
- What changes when motion enters the scene?
- What changes when the surface is different?
- What does the model learn? What does it fail to see? What does it only appear to understand?
Conditions under which machine perception can be trained. Tested. Observed. And made more accountable to the world it is meant to see.