Jon Barron

A thread of thoughts on radiance fields, from my keynote at 3DV: Radiance fields have had 3 distinct generations. First was NeRF: just posenc and a tiny MLP. This was slow to train but worked really well, and it was unusually compressed --- The NeRF was smaller than the images.

It makes sense that radiance fields trended towards speed --- real-time performance is paramount in 3D graphics. But what we've seen in AI suggests that magical things can happen if you forgo speed and embrace compression. What else is in that lower left corner of this graph?

A thread of thoughts on radiance fields, from my keynote at 3DV: Radiance fields have had 3 distinct generations. First was NeRF: just posenc and a tiny MLP. This was slow to train but worked really well, and it was unusually compressed --- The NeRF was smaller than the images.

In fact, it's the *opposite* of what we saw in object recognition. There we started with shallow+fast+big models like mixtures of Gaussians on color, then moved to more compact and hierarchical models using trees and features, and finally to highly compressed CNNs and VITs.

A thread of thoughts on radiance fields, from my keynote at 3DV: Radiance fields have had 3 distinct generations. First was NeRF: just posenc and a tiny MLP. This was slow to train but worked really well, and it was unusually compressed --- The NeRF was smaller than the images.

Generation three swapped out those voxel grids for a bag of particles, with 3DGS getting the most adoption (shout out to 2021's pulsar though). These models are larger than grids, and can be tricky to optimize, but the upside for rendering speed is so huge that it's worth it.

A thread of thoughts on radiance fields, from my keynote at 3DV: Radiance fields have had 3 distinct generations. First was NeRF: just posenc and a tiny MLP. This was slow to train but worked really well, and it was unusually compressed --- The NeRF was smaller than the images.

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

There's more good stuff in the paper: arxiv.org/abs/2502.10647. This paper is maybe a little strange because I "wrote it for the LLMs" --- it's a deep dive into a low level model component that would normally be spread across the supplements of several papers. Not planning to submit it anywhere.

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

Implementing f() well is tricky --- the naive way is numerically unstable and expensive. But I found a nice way to decompose the math such that it's super stable and cheap to evaluate. The trick is just phrasing everything in terms of expm1 and log1p while avoiding branching.

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

Now the cool stuff. f(x, λ) is defined for x>=0, but you can generalize it x<0 inputs by giving them their own λ. Changing (λ-, λ+) values lets you interpolate between sigmoid-/logit-/exp-/log-shaped functions. This family also *exactly* reproduces the ELU activation function!

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

f'(0.5 * x^2) yields a long list of kernel functions from the literature: rational quadratic kernels, the inverse kernel, the Gaussian/RBF kernel, the quadratic kernel, the multiquadric kernel, the inverse multiquadric kernel, and a non-normalized Student's T-distribution.

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

If you use those losses as negative log-likelihoods you get a family of PDFs that interpolates between Cauchy, Gaussian, and Epanechnikov distributions. This is again a superset of my CVPR2019 paper, where the new distributions (λ>1) have a finite domain that narrows as λ grows.

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

Power transforms by themselves are a really useful tool for "curving" data, as any gray-haired statistician will attest. I've used a half-finished version of this one in the past for curving metric distance along rays in Zip-NeRF: x.com/jon_barron/s...

I just pushed a new paper to arXiv. I realized that a lot of my previous work on robust losses and nerf-y things was dancing around something simpler: a slight tweak to the classic Box-Cox power transform that makes it much more useful and stable. It's this f(x, λ) here:

Here's an explainer video of my first paper in this line, before I understood things fully: www.youtube.com/watch?v=BmNK.... This new paper (arxiv.org/abs/2502.10647) broadens the same idea to a wider range of things: curves, losses, kernels, PDFs, bumps, and activation functions.

My pitch for an "LLM-native" alternative to citation count/h-index etc: 1) Train an LLM on a new paper and record the average loss during training. 2) Evaluate the retrained LLM on benchmarks. Your Google Scholar records avg_train_loss * benchmark_delta, and you go write your next paper.

fun test for image and video generation systems: add "the camera is upside down" to the prompt (especially for shots of people) and then vertically mirror the output. Even the best models struggle, with upside-down teeth and blinks, and gravity tugging everything up. Here's Veo 2.

I fed the "spinning dancer" illusion (a silhouette of a spinning figure that can be seen as rotating clockwise or counter-clockwise, left) into Runway Gen-3 (right). It resolved the ambiguity by having the dancer face the camera and oscillate, which is kinda clever.

`An advertisement for a shoe made of broccoli romanesco, rotating on a turntable` #Veo2