iiinova — where your digital self keeps evolving

Stefan Gugler⚗️化学 @stevain.bsky.social · 28/06/2026

A 'side project' of mine just got published in ChemTexts! (linked in last post) 1/3 I always really loved etymology and figuring out where certain words come from. The names of chemical elements are particularly interesting, as meaning, phonetics and societal stuff gets mashed together.

decomposition of the Chinese character for lithium into a semantic compound that means 'metal' and a phonetic compound that sound like 'li'

222

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

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)

100

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

5/ Molecules in 8 Steps: VP-ISSNR achieves 74% stability with 8 steps, 95% with 64 (SDE). Beats all baselines!

100

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

4/ We propose a new VP schedule 📈: Exponential inverse sigmoid SNR (ISSNR)→ rapid decay at start/end. Generalizes Optimal Transport Flow Matching.

110

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

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).

100

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

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)

100

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

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!

100

Stefan Gugler⚗️化学 @stevain.bsky.social · 12/03/2025

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. 🧵

151

Stefan Gugler⚗️化学 @stevain.bsky.social · 31/01/2025

our GPU cluster tonight after the ICML deadline

040

Stefan Gugler⚗️化学 @stevain.bsky.social · 17/12/2024

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?

100

Stefan Gugler⚗️化学 @stevain.bsky.social · 19/11/2024

The Pillar of Autumn (i know, i know)

020

Stefan Gugler⚗️化学 @stevain.bsky.social · 11/11/2023

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

000

Stefan Gugler⚗️化学 @stevain.bsky.social · 13/09/2023

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.

110

Stefan Gugler⚗️化学 @stevain.bsky.social · 13/09/2023

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

121