Table of Contents Show
Workshop brief: build a place, not a postcard
For Jaakko Saari, Yoyogi Station was never an anonymous piece of Tokyo infrastructure. He passed through it regularly while working at a small game company, and the view of the sky between its steel arches stayed with him. Years later, that familiar place became the setting for a quiet film scene in which a woman waits for a train before leaving her old life. Sonagraf Oy therefore had to build more than recognisable railway architecture: the environment had to carry memory, departure and the slightly dreamlike quality the real station can acquire when it briefly falls silent.
The result is a controllable realtime reconstruction covering four tracks across the Yamanote, Chuo and Sobu sides. It combines an early-2000s setting with deliberate Showa-era character, was designed for an LED volume and remains usable in VR. Sonagraf also used it to test how close a fully realtime Unreal Engine scene using Lumen could come to the credibility of path tracing. In VR, the reconstruction becomes something like a digital place-memory: not merely a background, but somewhere a visitor can enter.
The exercise is not to copy every visible object until the storage array asks for representation. It is to identify what the cameras, performance target and interaction actually require, then build those parts at the accuracy the audience can perceive. Yoyogi exposes nearly every awkward decision in location work: incomplete plans, changing architecture, close cameras, repeated steel, emissive signs, wet materials, translucent gates, period dressing, passenger behaviour and a render budget that remains stubbornly finite. Sonagraf’s method transfers to buildings, exhibitions, showrooms, historical sites, training simulations and game environments.
Step one: define the deliverable and cameras
Project decision: the location had to support an emotionally loaded waiting scene, not a general tour of the station. That required the platforms and their large structural forms, but those forms also had to be removable when they blocked a camera. An additional shot from inside a train justified an early Yamanote carriage interior. Tracks were rarely visible. An underground passage, stairs and even carefully modelled restroom geometry had rather less justification. The restrooms never appeared in the film. Like many expensive lessons, they were very accurate.
Problem and solution: shots change while detailed assets are expensive. Sonagraf therefore built a rough scene first, checked camera height, lenses, sightlines, walking routes and possible interactive areas, then divided the environment into close hero geometry, modular structures, removable obstructions and areas that could be excluded. The result was a scene that could change with the film rather than hold it hostage. The transferable rule is simple: approve the space through the intended cameras before approving the detail.
Accuracy followed perception and purpose. Large Yoyogi elements were targeted to roughly five to ten centimetres because errors beyond about 20 centimetres began to feel wrong. Tactile paving required millimetre-scale attention because it sits close to both feet and camera. For simulation work, Sonagraf may target measurement error below one millimetre and use LiDAR when the job requires it. A film background, safety simulation and interactive showroom can share tools without sharing the same tolerance.
Step two: construct a reference system
Problem: no public railway blueprints existed, and today’s Yoyogi is not the station of the early 2000s. Sonagraf combined extensive photography, contemporary and historical video, Google Maps measurements, satellite imagery and direct observation. The reference set recorded not only geometry and scale but material response, construction logic, ageing, period and human use. Earlier footage was essential whenever renovation had erased the version required by the story.
That work was not always glamorous. Saari describes himself as occasionally looking like “a suspicious foreigner inspecting handrail paint with an iPhone flashlight”. He also photographed rail detail from a nearby crossing with suitable caution. No CG art is worth dying for, particularly ballast. Vending machines, plastic seats, advertisements, drains, chipped paint, pipe layouts and track hardware often establish place more efficiently than another generic concrete texture.
Result: photographs corrected what daily familiarity had distorted. Saari had remembered tactile blocks too large and misjudged walls, arches and seats. Japanese station chairs are small and tightly grouped; “at a Finnish station, there would probably be a three-metre gap,” he jokes. Memory remains a remarkably confident source of incorrect measurements. The rule is to treat recollection as a clue, not evidence, and to choose human scale references that fit the local population rather than appointing one generic mannequin as the ISO standard for humanity.
Step three: block out before adding bolts
The Yoyogi blockout began in Autodesk Maya and continued in Blender. Scale came from the iPhone Measure app and comparisons against photographed passengers. Boxes and planes represented platforms, walls, roofs and equipment. Sonagraf assembled those modules in Unreal, tested them against the approved cameras and replaced them in place with high-resolution versions.
Decision: geometry was justified by the closest plausible view, not by habit. Steel arches, roof components, bolts and tactile paving could not hide behind normal maps. Plastic chair units required unique geometry and materials. Stairs stayed because another shot might use them—“what are stairs but tracks for humans?” Dense construction was planned around Nanite from the beginning, allowing detail to survive close inspection without reflexively applying pre-Nanite polygon rules.
Roughly 90 per cent of the modelling was completed in Blender, with Maya retained in the wider pipeline for rigging and animation. Platforms, pipes, walls and other repeated structures used modular components. Variants could be assembled in engine instead of exported as a parade of nearly identical meshes. The result was faster iteration and a cleaner route from blockout to finished asset.
Step four: make import boring and predictable
Almost all geometry moved through FBX with mostly default import settings. Static meshes remained separate rather than being combined on import. Names followed Epic’s conventions with practical texture exceptions such as T_Stationplatform_ORM.png. Predictable names and separation made later replacement, profiling and automation possible. A clever folder structure introduced after the crisis is still a folder structure introduced after the crisis.
ORM channel-packed textures came from Substance 3D Designer through a custom node setup or from Substance 3D Painter through a customised template based on the texture-set name. Project assets lived below /YoyogiStation, with folders including /Props and /Structures. Marketplace assets retained their default folders so their internal relationships did not break.
The environment remained in one main level. Packed Level Actors and Nanite Level Instances supplied instancing and organisation, while repeated assemblies could still be collapsed when an individual element needed adjustment. This scene structure later delivered the project’s largest optimisation gain. Sonagraf’s conclusion is worth carrying into any large environment: organisation is not administrative tidiness; it is performance engineering performed early enough to remain useful.
Step five: turn repetition into reusable controls
The project contains little game logic, but most repeated assemblies became Blueprints. Sonagraf kept the systems deliberately modest: sign-and-light combinations, movable platform gates, adjustable water, structural dirt and reusable packed assemblies. The immediate result was faster iteration; the longer-term result was a library of production controls that shortened later lighting variations, client revisions and location adaptations.
Platform gates use an Event Blueprint to control opening, animation and speed. A CRT-style platform monitor can be added later. Floor puddles use Absolute World Position in the material editor, mixing textures at different scales to conceal tiling. Material Instance Parameters expose water quantity and weathering, allowing the scene to move between dry daylight and a horror director’s preferred municipal neglect without repainting every surface.
Reusable systems also require maintenance. Rapid engine changes and incomplete documentation can complicate stable media pipelines. Saari remains enthusiastic about Blueprints while watching the transition towards Unreal Engine 6 and the long-term position of visual scripting. The transferable lesson is not panic: version critical tools, document their inputs and outputs, and freeze dependable production configurations when a show cannot absorb surprises.
Step six: light for the renderer you have
Early tests used Ultra Dynamic Sky, skylight, directional sunlight and Lumen global illumination. The first production version ran in Unreal Engine 5.3. Lumen reflections provided the base, while hardware ray tracing increased detail and supplied ray-traced shadows. Exposure remained fixed and camera-controlled. Sonagraf increased indirect-lighting intensity while managing exposure to keep daylight readable.
Problem: small emissive signs produced conspicuous Lumen noise. Solution: Sonagraf separated appearance from illumination. The visible glowing element became translucent and stopped contributing emissive light, while a matched rectangular light handled the actual illumination. A Blueprint linked the surface, light and shared controls. Result: the signs were cleaner, easier to art-direct and reusable across the station. It also avoided adjusting every sign individually, a task normally reserved for people being punished by production management.
The daytime scene omits fluorescent tubes that the camera never sees. Ultra Dynamic Sky allows a quick night variation, and signs can read its parameters. Dynamic fluorescent fixtures would require only a few more hours. The larger rule is to separate the permanent location from controllable presentation states: daylight, night, emergency, training and client-review versions can then share one structure.
Step seven: art-direct the evidence of use
Most surfaces passed through Painter with ambient-occlusion dirt baked from low-poly sources. Additional roughness variation came from Unreal material nodes, while decals supplied leaks, stains and black chewing-gum marks. Fixed camera exposure controlled colour and contrast. An iPhone served as a handheld virtual camera; an iPad on a SmallRig shoulder rig provides another physical interface. These tools kept art direction connected to the way the final camera would experience the station.
The first versions were too clean. Tokyo humidity, the rainy season, dirty pipes, rust and water are part of Yoyogi; its roofs are surprisingly poor at keeping out every drop. The final design blends older and newer structural forms to retain the desired Showa character within the early-2000s setting. Saari’s affection for the steel arches and the strips of sky between them guided what received attention. Films by Yasujirō Ozu and other Showa-era references guided the mood. Faithfulness therefore meant reconstructing the period and emotional truth required by the scene, not blindly copying today’s station.
Cultural accuracy is not decoration. Saari points out that international productions still flatten Japanese, Chinese and Korean conventions into a generic East Asian shorthand. For Sonagraf, passenger spacing, posture, queueing, dress and the way people wait are part of the reconstruction. More than ten passenger archetypes are planned with MetaHuman, custom clothing from Marvelous Designer and motion capture from a Rokoko suit. A technically accurate platform full of culturally implausible passengers is merely a different uncanny valley.
Step eight: profile the scene, not the folklore
The original target was 60 fps with headroom on an LED volume, with the same environment also working in VR. Sonagraf’s internal measurements put favourable Unreal Engine 5.3 runs above 90 fps, while LED-volume tests at Epic settings dipped slightly below 60 and peaked comfortably above 90. Later project builds were reported at roughly 120 fps in realtime and close to 200 fps with baked lighting on a suitable desktop. These project-specific figures were not independently reproduced at press time.
The stated hardware range is realtime output around 2K resolution on a modern gaming laptop and 4K on a capable desktop, with DLSS and supersampling available. Exact performance depends on GPU, engine build, drivers, resolution, LED processor and content configuration. Test that complete delivery chain before production. A frame-rate anecdote is not a delivery specification.
Measure with GPU Visualizer and stat fps, stat unit, stat gpu, stat rhi and stat memory. Inspect Shader Complexity, Light Complexity, Nanite and Lumen visualisations. Monitor streaming virtual textures and texture memory. Use stat unitgraph to expose CPU and GPU frame-time problems. Check light attenuation, reflection rays, shadow invalidation, translucency and overdraw before touching geometry because somebody on a forum once had a difficult shrub.
Step nine: optimise what the measurements identify
Measured priorities: Packed Level Actors produced the largest improvement by turning repeated structures into instances. Lowering in-engine texture resolution from 4K to 2K came second; a 4K texture contains four times the pixels of a 2K texture, and PBR sets repeat that expense across several channels. Reducing dedicated reflection rays through r.Roughness.Min and related maximum controls came third, especially on lower-end ray-tracing hardware.
Result: the team measured combined gains of around 20 to 30 per cent in some scenarios with negligible visible change, particularly under camera movement. Virtual Shadow Maps suited the Nanite-heavy environment. Light attenuation radii stayed tight, with indirect intensity used when a brighter contribution was needed. Platform-door glass was simplified because glass seen through glass compounded transparency cost. Extra props were merged or instanced, while future passenger shots may divide the environment into zones.
Traditional polygon advice did not always survive testing. Track units exceeded 300,000 triangles because ballast used displaced geometry. Reducing them to roughly 50,000 triangles produced no measurable gain. A dense Nanite version was lighter than a low-poly version carrying a 4K normal map because Nanite geometry compressed well while the normal map did not. Removing unseen faces and merging already-simple walls also offered little value; watertight meshes could perform better.
Geometry is not universally free. Old foliage made from large translucent cutout cards can still demolish performance through overdraw, while real Nanite leaf geometry may behave better. The transferable rule is to optimise for the active renderer and the pixels that dominate the frame. Textures, reflection rays, shadows, translucency, material complexity and overdraw remain the habitual offenders. This profiling work has also become a Sonagraf service: the company audits and rescues Unreal projects that already look acceptable but cannot meet their delivery frame rate.
Step ten: turn the project into a reusable pipeline
If rebuilding the station now, Sonagraf would preserve instancing more deliberately between Blender and Unreal with Houdini and Unreal’s Procedural Content Generation framework. The company now uses that combination for similar work and reports significant pipeline improvements. The lesson is not that Yoyogi was built incorrectly; it is that a good project should leave the next one with better repeated structures, materials, controls and validation steps.
The same systems can support another station, exhibition, showroom, interactive environment or historical reconstruction. A Helsinki railway environment would still demand local measurement, materials and cultural observation, but it would not be Sonagraf’s first station. The same reading of incomplete source material has already travelled beyond architecture: Sonagraf reconstructed a high-end tugboat from shipbuilders’ two-dimensional drawings, including plans still produced by hand. Reusable production knowledge is the useful output hiding behind the pretty render.
The experience behind the workshop
The method comes from Sonagraf Oy, an XR and interactive-media company operating between Finland and Japan. It was established in Japan in 2022 and in Finland in 2026. Head of production Jaakko Saari has 15 years of experience as a CG generalist and worked as a technical artist at Adobe in Japan in 2019 and 2020. Sonagraf is run by Saari as head of production and CEO Sachiko, who manages the company’s operations; additional specialists join projects as required.
Saari teaches realtime graphics at Digital Hollywood University in Tokyo and provides corporate Unreal Engine training for Japanese media companies. He moved from Finland to Japan in 2006, first worked as a photographer and later graduated from the New York Institute of Photography. That combination of image-making, technical production and more than 20 years of local experience explains the workshop’s central concern: a place is not credible because its dimensions are plausible; it is credible when structure, light, wear, behaviour and cultural context agree.
Working in English, Finnish and Japanese across Europe and APAC, Sonagraf focuses on Unreal environments, real-location reconstruction, XR, simulation, Blueprint development, optimisation, virtual-production assets and on-site capture. Its pipeline combines Unreal, Blender, Houdini, Maya and Adobe Substance 3D with HDRI photography, 3D scanning and measured material capture. The company can begin with photographs, plans, scans or existing client assets and build the missing links into a production-ready environment.
The choice of capture or modelling method remains practical. Sonagraf favours traditional modelling where editable lighting, topology and UVs matter more than the speed of a photogrammetry or Gaussian-splatting capture, while using scans when their accuracy serves the brief. Measured PBR references from physicallybased.info complement the company’s own material captures. The rule throughout the Yoyogi project applies here as well: choose the representation that the final production can control.
Yoyogi Station began as the backdrop for one film scene. It became a test of realtime photorealism, a reusable production system and a record of a place Saari had known for years. It also marks his move from building individual assets to supervising the decisions that make an entire environment work. That combination is Sonagraf’s strongest proposition: not simply modelling what a location looks like, but understanding what must be measured, controlled, animated, optimised and preserved so the digital version can do more than the original brief required.
Saari’s next personal environment is a public sento bath for the same film, with a Japanese shrine also on the list. Sonagraf is available for projects in Europe and the APAC region. Contact jaakko@sonagraf.com, or view additional work on ArtStation.
https://sonagraf.com/
https://www.artstation.com/yaschan
https://www.linkedin.com/in/jaakko-saari-crw/
