Next Event Estimation clearly wins - fewer samples, faster convergence, same physical accuracy ⚡

However, it seems like NEE results in slightly brighter images. That points to a bug, the converged outcome should be identical.

A proper denoising algorithm could speed up image generation further.

- N = 1000
Both images finally look similar — but NEE reached this quality already with much fewer samples 🙌. So those extra samples were maybe not needed.

Render time: ~10–15 minutes 💻

- N = 100 samples
NEE is almost fully converged, while path tracing still has plenty of grain and noise

- N = 10 samples
At 10 samples, NEE is starting to clean up nicely 👌

Path tracing still struggles — lots of missed light paths and black areas

- N=1 sample:
With just one sample per pixel, brute force path tracing wastes tons of rays — you can see all the black pixels 🕳️

NEE, on the other hand, already captures the main light contributions 💡 (still noisy, of course)

⚙️ Method: Ray tracing
🎥 1 sample per frame, 30 fps
⏱️ ~1–2 minutes render time / video

🌍🌕 Switching perspectives to a geocentric scene with moon and stars in relative motion.

Don't get dizzy...

How it's going. 🧊

#debugging

⚙️ Making it faster
Path tracing gives the best result (check out e.g. the ceiling), but is super slow.
To optimize, I’m adding:
• Direct light sampling — directly sampling visible light sources to reduce wasted rays.
• Denoising — cleaning up noise so fewer samples are needed for a smooth result.

🌌 Path Tracing Renderer
It simulates full global illumination: Light bounces multiple times, capturing indirect light and color bleeding.
The single-sample image is almost black, most rays never hit the lamp.
After ~15000 samples/pixel (≈3 hr), the scene is still noisy.

🔦 Ray Tracing Renderer
Introduces stochastic light sampling for reflections, refractions, and light sources.

The first image shows a single sample per pixel (noisy but fast). The second is the converged result after 2500 samples.

I have to figure out what causes the few "hot pixels"...

💡 Deterministic Renderer
Adds physically based light interactions: direct light (and therefore shadows), reflections, and refractions are handled explicitly.

No randomness — everything is computed deterministically for stable, noise-free results.

🎨 Simple Renderer
This one ignores all light transport — it just shows the raw base colors of materials.
It’s useful for debugging geometry, textures, and UV mapping before diving into lighting.

Performance-wise, it is not so great.

While having already parallelized my code, rendering each of these images still takes around 7-8 minutes, i.e. 0.002 fps. :(

My next goal is to speed up my renderer significantly. I already have two promising ideas. Let's see.

With this new material module, I will be able to simulate a bunch of interesting scenes and render lots of images.

But first, I might increase the fidelity of my generated images by simulating the Fresnel effect. But more on that next time.

Finally, through all these properties that I can adjust, I can try to model metallic surfaces.

Here, I tried to simulate a bronze, silver, and golden sphere.

Just like for reflections, the glass might have a rough structure which adds a bit more randomness to the refractive interaction.

You get a more "milky" glass.

Moving on to transparent materials and glass spheres. We can simulate refraction by increasing the refractive index of the spheres and using Snell's law.

Check out the light patterns on the floor.

Reflections are not always perfect. Mirroring surfaces might not be smooth. We can increase the surfaces roughness and see what happens to the reflected light.