Unreal Engine Technical Artist — Synthetic Human Face & Eye Data

About the work Senseye builds computer-vision models that measure fine-grained ocular and facial signals from ordinary smartphone video. To train those models we generate synthetic human face datasets in Unreal Engine — photoreal 3D heads where we control every variable and therefore know the exact ground truth: pupil diameter, iris diameter, gaze vector, head pose, eyelid aperture, blink state, lighting, camera geometry. That perfect ground truth is the whole point. Real video gives us approximate labels; synthetic gives us exact ones. We train on synthetic, augment with real, and infer on real. We have working results already — the bottleneck now is the quality and throughput of the synthetic asset pipeline itself. We are looking for a contractor who can close the domain gap. The closer our rendered faces are to real phone footage of real people, the better our models transfer. That is the job. What makes this role unusual Two things, and we need both: 1. Eyes are the product, not a detail. Most character work treats eyes as a small part of a face. For us the eye is the measurement surface. We need someone who will care about corneal refraction, limbal ring falloff, iris stroma depth and parallax, scleral vascularity, the tear meniscus and caruncle, eyelid occlusion and lash shadowing — and who can build a rig where pupil diameter is a first-class, physically calibrated parameter that drives geometry and shading correctly rather than a texture cheat. 2. The primary workflow is headless, not the Editor. We run Unreal largely from the command line. Our generation loop is scripted, batched, seeded, and reproducible — we need thousands of controlled variations, not a dozen hero renders. The Editor UI is a validation and authoring tool for us, not the production path. We also use AI coding agents (Claude Code, OpenCode, on-prem local LLMs) to write and iterate the Python/automation layer, and we want someone who works that way naturally rather than someone we have to convince. If you are a brilliant cinematic artist who works exclusively in the viewport, this is probably not a fit. If you are a technical artist who writes the tool and then runs it 10,000 times, read on. Scope of work 1. Photoreal, demographically representative head assets Build and maintain a parameterized library of MetaHuman-based (or equivalent) heads covering a genuine cross-section of the world: skin tones across the full melanin/ITA range, eyelid and periorbital morphology across ancestries (including epicanthic fold variation), iris colors from light blue through very dark brown, age range, facial hair, and eyewear. Dark irises are a specific, called-out challenge. Low pupil/iris boundary contrast is where our models struggle most and where synthetic realism matters most. We need the rendering to hold up there. Correct skin shading — subsurface scattering, specular breakup, pore and micro-normal detail, sweat/oil variation — under the lighting conditions phones actually see. Grooms: brows, lashes (a real occluder for eye measurement), scalp hair. Lash density and geometry need to be varied, not one preset. 2. Eye and face rig with physically meaningful controls Pupil diameter driven in millimeters, with correct iris compression/dilation behavior and consistent geometry-to-shading coupling. Gaze aimed by explicit vector or target, with realistic ocular kinematics — saccades, smooth pursuit, vergence, microsaccadic jitter, vestibulo-ocular coupling with head motion. Blink and lid-aperture control with realistic partial blinks and lid-lag, not binary open/closed. Facial expression driven by FACS-style action units or an equivalent parameterization we can sample programmatically. Head pose in 6DoF with plausible natural motion. 3. Headless generation pipeline Command-line Unreal execution (UnrealEditor-Cmd, -ExecutePythonScript, Movie Render Graph / MRQ from CLI) on our own hardware, running unattended and in parallel. Linux support matters to us. Our Unreal Engine 5.8 pipeline and batch infrastructure are Linux-first. Windows-only pipelines are a significant limitation. Python and Blueprint scripting for batch editing, assembly, sculpting, conforming, wardrobe, rigging, and texture operations — MetaHuman 5.8 exposes this scripting surface and we intend to use it heavily. Seeded, deterministic randomization. The same config must produce the same dataset. Every render carries a manifest: engine version, asset versions, all sampled parameters. Sensible throughput and cost per frame. Tell us where you'd spend render budget and where you'd cheat. 4. Ground-truth export Per-frame, machine-readable, in a schema we agree on up front: Pupil diameter (mm) and iris diameter (mm) Gaze direction in camera and world space; point of regard Head pose (6DoF), eye pose relative to head Eyelid aperture, blink state, occlusion fraction of iris/pupil 2D landmark projections for face, lid margin, limbus, pupil boundary Segmentation masks (iris, pupil, sclera, lids, skin, hair) via custom stencil / object ID passes Depth pass, camera intrinsics and extrinsics, lighting parameters Anything derivable from the scene that a labeler could not produce by hand — that's the value 5. Domain-gap engineering This is where we most want your judgment. Our renders must look like phone footage, not like a game cutscene: Front-facing smartphone camera simulation: realistic FOV, lens distortion, sensor noise model, rolling shutter, auto-exposure and auto-white-balance behavior, motion blur, and codec/compression artifacts. Lighting that reflects real capture conditions — indoor mixed color temperature, window light, overhead office fluorescents, screen glow from the phone itself illuminating the face, low light, harsh backlight. Corneal specular reflections and catchlights that behave correctly, since they're a real confound in ocular measurement. A point of view on how to measure the domain gap, not just assert it. 6. Documentation and handoff Our engineers will run and extend this pipeline. Code, conventions, and asset structure need to be legible to AI engineers who are not Unreal specialists. Requirements Deep, demonstrable Unreal Engine 5.8 expertise. Materials, shading, lighting, rendering, Sequencer, Movie Render Graph, Control Rig. MetaHuman fluency — Creator, Animator, DNA/rig manipulation, Groom, and the scripting APIs around them. Character/human realism as a specialty. Portfolio must show photoreal human faces you personally made, ideally with close-up eye work. Python scripting and command-line Unreal automation. You should be comfortable running the engine as a headless batch job on a Linux box. Comfort working with AI coding agents (Claude Code, OpenCode, Cursor, local/on-prem LLMs) as part of your normal iteration loop. Available for meaningful US-Central-overlapping hours and able to work on our infrastructure under our data-handling constraints. US persons only — this work touches a regulated data environment and stays on-prem; no third-party cloud services, no external model APIs. Able to engage as a 1099 independent contractor. This role cannot be brought onto payroll. You should be set up to invoice us directly, whether as a sole proprietor or through your own entity. Human-subjects research training may be required. If the work requires access to clinical subject data or videos for comparison or validation, the contractor will need to complete the applicable human-subjects research training/certification before receiving access. Senseye will cover the certification cost and compensate the contractor for the time required to complete it. Nice to have Prior synthetic data generation for computer vision or ML training (Unreal, Unity Perception, NVIDIA Omniverse Replicator, Blender/Houdini pipelines). Working knowledge of eye tracking, pupillometry, oculomotor behavior, or ophthalmic anatomy. Experience closing sim-to-real domain gaps and evaluating them quantitatively. Photogrammetry, 4D facial scan processing, or scan-to-rig workflows. Houdini for procedural grooms/assets; Maya/XGen. Render farm / distributed rendering orchestration. Experience in regulated, health, or research environments.