💻 computer science

Texel Splatting: Perspective-Stable 3D Pixel Art

Texel Splatting achieves perspective-stable 3D pixel art by rendering scene geometry into a world-space cubemap and splatting texels as world-space quads, thereby ensuring rotation and translation invariance without relying on problematic camera grid-snapping.

Original authors: Dylan Ebert

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dylan Ebert

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a 3D video game world, but you want it to look like retro pixel art (like the old Nintendo games). The problem is, in real life, when you walk around a room, things closer to you move faster across your vision than things far away. This causes a glitch in pixel art: the pixels start to "shimmer" or "creep" around, breaking the illusion of a solid, blocky world.

This paper introduces a clever new trick called Texel Splatting to fix that shimmer without having to freeze the camera in place.

Here is the breakdown using simple analogies:

1. The Problem: The "Shimmering Floor"

Think of a standard 3D game like a movie camera. If you move the camera, the pixels on the screen shift.

In Orthographic (Flat) View: Imagine looking at a wall through a window. If you slide the window left, the whole wall slides left by the same amount. You can just "snap" the pixels to a grid, and they stay perfect.
In Perspective (Real Life): Imagine looking down a long hallway. If you move your head, the tiles on the floor near your feet zoom past you, but the tiles at the end of the hall barely move. You can't "snap" the whole image to a grid because the near stuff and far stuff are moving at different speeds. This causes the shimmer.

2. The Solution: The "Magic Cube" (Cubemap)

Instead of trying to fix the pixels on the screen after the camera moves, Texel Splatting changes where the image is created.

Imagine you are standing in the center of a giant, invisible glass cube.

The Probe: You place a "camera" (called a probe) at a specific spot in the world. This camera takes a picture of the entire room in all six directions (up, down, left, right, front, back) and paints it onto the inside walls of that glass cube.
The Grid: To keep things stable, this "camera" doesn't float freely. It is magnetically snapped to a giant, invisible grid on the floor. As long as you walk within one square of the grid, the camera stays in the exact same spot.

3. The Magic Trick: "Splatting"

Now, here is the cool part. Instead of looking through the camera to see the scene, the computer takes every single little square (texel) painted on that glass cube and turns it into a floating 3D billboard (a quad) in the real world.

The Analogy: Imagine the glass cube is a set of stamps. The computer takes those stamps and slaps them onto the 3D world as physical, flat pieces of paper.
Why it works: Because these "papers" are attached to the world (the geometry), not the camera, they don't jitter.
- If you walk around, the camera moves, but the "papers" stay glued to the walls and floors.
- Since the papers were painted from a fixed spot on the grid, they don't shimmer. They just look like solid, stable pixel art blocks.

4. Handling the Blind Spots (The "Eye" vs. The "Grid")

There is one catch: If your "Grid Camera" is stuck in one spot, it can't see around corners or behind objects that are far away. If you walk past a wall, the Grid Camera can't see what's behind it.

The Fix: The system uses two cameras at once.
1. The Grid Camera: Stays snapped to the grid. It paints the stable, non-shimmering parts of the world.
2. The Eye Camera: Moves with you. It sees the new stuff (what's behind the wall) but it does shimmer a little.
The Blend: The computer mixes them. The stable Grid Camera does 99% of the work. The Eye Camera only fills in the tiny gaps of new things you just walked into. Because the "new" stuff is usually small, the shimmer is barely noticeable.

5. Why This is a Big Deal

Old Way: To get stable pixel art, you had to force the player to move in "snaps" (like a chess piece), which feels clunky and unnatural.
New Way: You can move the camera smoothly and freely, like a real person, but the world still looks like a perfect, stable 8-bit game.
Performance: It's actually very fast. Because the "painting" happens on the fixed grid cubes, the computer can reuse the work from the previous frame instead of recalculating everything every single millisecond.

Summary

Texel Splatting is like taking a photo of a room from a fixed spot, cutting that photo into thousands of tiny 3D stickers, and sticking them onto the real world. As you walk around, the stickers stay put, keeping the pixel art crisp and steady, while a second "helper" camera quietly fills in any new details you discover.

Based on the paper "Texel Splatting: Perspective-Stable 3D Pixel Art" by Dylan Ebert, here is a detailed technical summary covering the problem, methodology, contributions, results, and significance.

1. Problem Statement

The core challenge addressed is rendering 3D scenes in a pixel art aesthetic that remains stable under perspective projection.

The "Pixel Creep" Issue: In orthographic projection, camera movement shifts all pixels uniformly, allowing a simple "grid snap" to correct the image and keep pixels locked to geometry. However, perspective projection causes objects at different depths to shift at different rates on the screen.
Failure of Existing Methods: No single screen-space grid snap can correct the varying displacements caused by perspective. Existing solutions (like ProPixelizer) admit that pixel creep is unsolvable for perspective views using standard screen-space techniques.
Limitations of Prior Work: Previous attempts like "Texel Marching" use cubemaps but rely on screen-space raymarching, which can be computationally expensive or complex to implement efficiently.

2. Methodology

Texel Splatting solves this by decoupling texel stability from camera movement through a fixed-world probe system and world-space rendering. The pipeline operates in three main stages per frame:

A. Cubemap Capture (The Probe)

Fixed Origin: Instead of rendering from the camera, scene geometry is rasterized into a cubemap from a fixed "probe origin" in the world.
Grid Snapping: The probe origin is the camera position snapped to the nearest vertex of a world grid.
- Translation Invariance: As long as the camera stays within a grid cell, the origin remains fixed, ensuring the cubemap content does not shift.
- Rotation Invariance: Cubemaps are indexed by direction, not screen position, making them inherently rotation-invariant.
Data Capture: The cubemap stores position, normal, material, and object ID per texel.

B. Shading (Compute Pass)

Camera-Independent Lighting: Shading occurs in cubemap space, not screen space. This includes directional/point lighting and shadow ray tracing against scene geometry.
Aesthetic Processing:
- Posterization: Lightness is quantized into discrete bands using OKLab color space to create flat color steps.
- Outlines: Silhouette edges (object ID discontinuities) and creases (normal discontinuities) are detected and applied as lightness shifts.
Stability: Because shading happens in the fixed cubemap space, the visual output is stable regardless of camera movement.

C. Splatting (World-Space Quads)

Projection: Each visible cubemap texel is rendered to the screen as a world-space quad.
Depth Calculation:
- Each texel stores a scalar depth $d$ (Chebyshev distance from the probe).
- The world position $p$ is calculated as: $p = o + \frac{d}{\|r\|_\infty}r$ , where $o$ is the probe origin and $r$ is the direction vector.
- Quad corners are evaluated at $(u \pm h, v \pm h)$ , sharing the same depth value to avoid per-corner intersection tests.
Gap & Z-Fighting Handling:
- Quads are expanded beyond texel boundaries to fill gaps.
- To prevent Z-fighting (flickering) where adjacent texels have similar depths, expansion is constrained based on the grazing angle between the view direction and the cubemap face.

D. Handling Disocclusion (The "Eye Probe")

The Trade-off: A fixed probe cannot see geometry occluded from its specific viewpoint. As the camera moves, new geometry appears (disocclusion).
Hybrid System: The system maintains three probes:
1. Grid Probe: The stable, snapped origin (provides the main stable image).
2. Previous Probe: Holds the old grid cell for blending.
3. Eye Probe: Located at the actual camera position (captures disoccluded content).
Blending: When the camera crosses a grid boundary, the system crossfades between the old and new grid probes using a Bayer dither pattern. The "Eye Probe" fills gaps where the grid probe cannot see, accepting that these specific regions may exhibit some shimmer (unstable texels) because the eye probe moves with the camera.

3. Key Contributions

Perspective-Stable Pixel Art: A novel technique that achieves stable pixel art under perspective projection, a problem previously considered unsolvable for general 3D scenes without orthographic constraints.
World-Space Splatting: Replaces screen-space raymarching (Texel Marching) with a more efficient world-space quad projection, leveraging the GPU's rasterizer for the final display.
Decoupled Stability: Demonstrates that stability can be achieved by fixing the rendering origin to a world grid rather than trying to correct the camera or screen space.
Hybrid Probe Architecture: Introduces a multi-probe system (Grid + Eye + Previous) to balance stability with the need to render disoccluded geometry.

4. Results & Performance

Performance:
- Mobile: Runs at 40 FPS on an iPhone 15 (A16 GPU) at 3842 texels per face.
- Desktop: Runs in <4 ms per frame on an RTX 4090.
Efficiency: The pixel art aesthetic allows for low-resolution cubemaps, keeping the shading budget small. The camera-independent shading allows for caching across frames.
Visual Quality: Successfully renders 3D scenes with flat shading, outlines, and posterized colors that remain locked to geometry during camera movement, eliminating the "shimmer" typical of perspective pixel art.

5. Significance

Redefining Pixel Art in 3D: This technique bridges the gap between 2D pixel art aesthetics and 3D perspective, enabling games and applications to use true 3D geometry without sacrificing the "locked" feel of 2D sprites.
Optimization for Mobile: The method is highly optimized for mobile hardware, making high-quality, stable pixel-art 3D rendering accessible on consumer devices.
Future Directions: The paper identifies disocclusion as the primary remaining limitation. Future work focuses on adaptive probe placement, handling animated geometry, and improving specular material rendering within the stable grid framework.

In summary, Texel Splatting offers a robust solution to the "pixel creep" problem by moving the rendering origin to a fixed world grid and projecting geometry as world-space quads, achieving a level of visual stability in perspective views that was previously unattainable.