A facial composite is a visual representation of a person’s face, typically created from an eyewitness description to help police identify an unknown suspect. You’ve almost certainly seen one: those slightly uncanny face images shared on the news or posted on bulletin boards during criminal investigations. Their primary purposes are to identify or eliminate a suspect and to corroborate a victim’s or witness’s account of events.
Composites are not meant to be perfect portraits. They function as investigative tools to narrow the field of possible suspects, not as a method of directly identifying a perpetrator. If the only evidence available comes from an eyewitness, generating suspects through a facial composite may be the only way to move an investigation forward.
How a Composite Gets Made
The traditional method involves a trained police artist drawing freehand based on a witness’s verbal description. The artist works through the suspect’s facial shape, individual features, and distinguishing marks like scars, tattoos, or unusual physical characteristics. This is a collaborative, back-and-forth process where the witness guides adjustments until the image looks as close to their memory as possible.
Starting in the 1980s and 1990s, software systems began replacing or supplementing sketch artists. These early digital tools, such as E-FIT and PRO-fit, took a feature-based approach. A witness would select individual components (a nose shape, an eye shape, a jawline) from a database, and the software would merge them into a single face. The process still relied heavily on detailed verbal descriptions of each feature.
Feature-Based vs. Holistic Systems
The feature-based approach has a fundamental problem: it doesn’t match how people actually remember faces. Most of us don’t store a mental catalog of someone’s nose, eyes, and mouth as separate items. We remember faces as whole impressions, which makes it surprisingly difficult to describe individual features in isolation.
Holistic composite systems were designed to work with this reality rather than against it. The most notable is EvoFIT, which uses an evolutionary approach. Instead of picking features one at a time, a witness views a screen of roughly 70 complete faces generated by the system. They select the ones that most resemble the suspect, and the software “breeds” those selections together, producing a new generation of faces that share characteristics of the ones chosen. The whole face changes with each iteration, not just a single feature, guided by underlying models of facial shape and texture. The witness repeats this process through several rounds until the composite converges on a likeness.
Research comparing the two approaches found that composites from older feature-based systems like PRO-fit were correctly identified about 18% of the time when shown head-on. EvoFIT has produced more identifiable composites than both E-FIT and PRO-fit, though overall naming rates remain relatively low across all systems. A composite that looks “sort of like” someone is often enough to generate a lead, even if it wouldn’t pass for a photograph.
Why Memory Makes Composites Difficult
The accuracy of any composite depends almost entirely on the quality of the witness’s memory, and human memory is far less reliable than most people assume. Stress, lighting conditions, the duration of the encounter, and the time elapsed before the composite session all affect how much detail a witness can provide. A witness who saw someone for three seconds during a robbery will produce a less useful composite than one who interacted with a suspect for several minutes.
There’s also a phenomenon researchers have studied where the act of describing a face in words can actually interfere with the visual memory itself. Translating a mental image into verbal descriptions (“the nose was medium-sized, slightly upturned”) can overwrite the more nuanced, holistic memory of what the person looked like. This is one reason holistic systems were developed: by letting witnesses respond to whole faces rather than describe features verbally, the process is designed to preserve the original memory more faithfully.
Creating a composite can also potentially affect a witness’s later ability to identify the actual suspect in a lineup. Newer holistic systems were specifically designed to minimize this interference, and research suggests they do not lead to more misidentifications compared to older feature-based systems.
Role in an Investigation
A composite serves a specific and limited role. Police release it publicly or circulate it internally to generate tips and narrow their suspect pool. If someone recognizes the face, that person becomes a lead to investigate further. If a suspect is eventually identified, the witness may then be asked to pick the perpetrator out of a lineup or video parade, which is a separate process entirely.
Composites are not considered definitive proof of identity. No reliable “litmus test” exists that allows law enforcement or courts to objectively measure how accurate a given composite is. Defense attorneys can file motions to suppress composite evidence, and the question of whether a composite should be shown to a jury remains contested. Some legal experts argue that the initial description a witness gives to police, such as during a 911 call, is often more accurate than a composite produced hours or days later after the memory has had time to shift.
How AI Is Changing the Process
Artificial intelligence is beginning to reshape how composites are created. Generative adversarial networks, a type of AI where one algorithm generates images while another evaluates how realistic they look, can produce highly detailed synthetic faces at resolutions up to 1024 by 1024 pixels. The two algorithms push each other through iterative feedback: the generator gets better at creating convincing faces, and the evaluator gets better at spotting flaws.
In the context of composites, this technology could allow systems to generate more photorealistic results from witness input, moving beyond the slightly artificial look that current composites often have. AI models can also create entirely synthetic face datasets for training purposes, which is useful for improving composite software without relying on real individuals’ photographs and the privacy concerns that come with them. Several variants of this technology have been developed to address early challenges like training instability, and recent models can generate faces with both anatomical accuracy and perceptual realism that makes them difficult to distinguish from real photos.
The core challenge remains the same regardless of the technology: a composite is only as good as the memory behind it. Better software can translate a witness’s recollection into a more realistic image, but it cannot recover details the witness never encoded in the first place.