7/ Why it works? Our empirical analysis shows:

1. Straight trajectories near data (t ≈ 0) are important (see in the inset plot)
2. Broad support of pₜ(𝐱) early on → robust to errors (note how SMLD goes from small to huge range instead of staying the same)

5/ Molecules in 8 Steps:

VP-ISSNR achieves 74% stability with 8 steps, 95% with 64 (SDE). Beats all baselines!

4/ We propose a new VP schedule 📈:

Exponential inverse sigmoid SNR (ISSNR)→ rapid decay at start/end. Generalizes Optimal Transport Flow Matching.

3/ VP variants improve existing schedules:

Take SMLD/EDM (exploding TV) → force TV=1. Result: +30% stability for molecules with 8 steps

(x-axis is NFE=number of function evals).

2/ Most schedules (like EDM by Karras or SMLD (Song & Ermon) let TV explode (VE=variance exploding).

We show constant TV (variance preserving, VP) + optimized SNR works better (ISSNR)!

(it's a wild table, sorry, but notice our VP variants I circled)

1/ Problem: Diffusion models are slow due to repeated evals but reducing steps hurts quality if the noise schedule isn’t optimal. Other schedules passively adjust variance. Can we do better?

🔑Insight: control Total Variance (TV) and signal-to-noise-ratio (SNR) independently!

We have a new paper on diffusion!📄

Faster diffusion models with total variance/signal-to-noise ratio disentanglement! ⚡️

Our new work shows how to generate stable molecules in sometimes as little 8 steps and match EDM’s image quality with a uniform time grid. 🧵

our GPU cluster tonight after the ICML deadline

stuff like this, (Ala)_2 at 2 fs or water at 1 fs? at <=0.5 fs they wouldn't explode for curl free forces, i assume?

The Pillar of Autumn (i know, i know)

its the dawn of the new millennium and you're somehow listening to panchiko

me rn

The issue is that Bayesians (me too!) are sometimes quick to doodle some Gaussians and smugly show how we update beliefs. But the distribution of the belief (e.g. fat-tailed or not??) doesn't often enter the convo.

Student-t (left) vs Gaussian (right) with their log probabilities (below), sort of illustrating the student's fat tails.