Context and objective
Metamorphosis is a Kafka-inspired first-person adventure in which the protagonist, turned into an insect, explores a world that is gigantic at his scale. The assignment was to bring it to Meta Quest 2 and Quest 3, and it was one of the most technically complex projects I have worked on.
I had never worked in VR before, so I had to learn many new techniques and concepts. On top of that came extremely tough performance requirements, hardware equivalent to a 2020 smartphone, a two-person team and a schedule of about nine months.
Team and responsibilities
I was in charge of the graphics optimizations: instancing, materials, LODs, draw distances, HLODs and lighting, among other tasks.
I was also one of the main designers of the conversion of the mechanics to VR, being the only one on the team with extensive experience as a VR player. The implementation of movement and climbing was mainly handled by my colleague; I collaborated on it and, above all, on reviewing the results. Some design decisions ended up changing during the process at the client’s request.
The hardware
The Quest 2 was the limiting platform. The goal was to run the game at its maximum resolution and 90 FPS, which leaves about 11 ms per frame to render the scene twice, once per eye. In VR this is not a soft target: dropping below it causes discomfort and motion sickness.
The original game had been designed for much more powerful platforms and was not built for this budget:
- Thousands of draw calls from unbatched meshes, saturating the render thread.
- Geometry with irregular and often unnecessarily high densities.
- Materials with many textures and expensive per-pixel effects.
- Open spaces that prevented effective culling.
We measured performance on the device itself with the developer tools Meta provides for Quest.

Quest 2 uses a Qualcomm Snapdragon XR2 chip, based on a mobile architecture.
Massive instancing
The most urgent problem was the render thread. Each independent static mesh means at least one draw call, and environments like the one in the image contained thousands of copies of the same brick, the same plank or the same decorative object.
The solution was instancing: grouping all copies of the same mesh with the same materials into a single hierarchical instanced static mesh (HISM) component, which the GPU draws with a single call. We used editor tools that automatically group a level’s actors by their mesh and material combination, together with HLOD generation for distant geometry.
In the most affected areas, draw calls fell from roughly 4,000 to just 200–400.

This area, packed with bricks and repeated objects, exceeded 4,000 draw calls.
Materials and per-pixel cost
In VR the resolution is very high and everything is rendered twice, so every per-pixel instruction counts. We did an exhaustive review of the game’s materials:
- Fewer textures per material where feasible.
- Removing distortion effects from shaders, which force the already-rendered image to be read again.
- Dynamic lights converted to static where the design allowed it, with lighting baked into lightmaps.
Why textures matter so much on a tile-based GPU
Mobile GPUs like the Quest’s Adreno use tile-based rendering: they split the screen into small regions and render each one in very fast on-chip memory, writing the result to main memory only once. This saves the bandwidth a desktop GPU spends continuously writing and reading the framebuffer.
Textures, however, live in main memory, which is shared with the CPU and has limited bandwidth. Every read that misses the texture cache has to be fetched from there, at a high cost in time and energy. Reducing the number and size of the textures a material uses improves the cache hit rate and eases that bandwidth pressure. The actual benefit depends on how each material uses them, but on such limited hardware it is one of the most cost-effective optimizations.
Open spaces
Indoors, walls and doors hide most of the scene and culling discards what is not visible. In open spaces, almost everything is potentially visible. To limit what was drawn at a distance, we combined several techniques:
- Very aggressive LODs at long distances.
- Shorter maximum draw distances for small objects, assigned with a studio tool that computes each mesh’s distance from the size of its bounds and an adjustable curve. This way a small object disappears much earlier than a large one, without configuring each one by hand.
- Exponential fog as a visibility limit.

Open areas reduced the effectiveness of culling.
Fog was especially useful in an underground village full of houses and characters. A classic exponential fog allowed us to impose a much harsher limit on draw distance without breaking the atmosphere of the place, which is already gloomy, mysterious and subterranean, with mushrooms growing everywhere. It is a good example of an optimization that leans on the art direction instead of fighting it.
VR adaptation
Besides the graphics, VR mechanics had to be introduced into gameplay designed for a traditional screen. This meant a radical change to the movement system:
- A climbing system that lets you climb practically any surface, as in the original game, but using your hands.
- Immersive movement based on the use of the hands, allowing some local physical movement.
My role here was mainly design and review: deciding how each mechanic should feel in VR, spotting comfort issues and validating the results with the team.
Result and final state
When we left the project, only two open areas remained where performance on Quest 2 fluctuated between 60 and 90 FPS: the crowded village and a gigantic room with a gramophone, full of characters and objects.
Due to various issues with the company that commissioned the project, our team had to leave development before completing it. Another studio finished the game and released it with hardly any changes from the state in which we left it: there were no Quest 3-specific graphics improvements, VR-specific accessibility options were missing and some graphics and performance issues remained unresolved. This case study describes our contribution and the state we reached, not the final result of that other team.
Even so, I am very proud of the technical level we achieved with such a small team and such a tight schedule.
What I learned
- In VR, the frame budget is non-negotiable, and it forces you to think about the cost of every pixel and every draw call from the very beginning.
- On mobile hardware, memory bandwidth is as important as compute power.
- The best optimizations are sometimes art decisions: a well-chosen fog can be worth more than weeks of LOD tuning.
Demonstration
First minutes of gameplay on Quest 3.
