Rayhan Zirvi
Portrait of Rayhan Zirvi

About

I am a senior at Caltech studying computer science. My research focuses on generative modeling, particularly on understanding the design choices behind diffusion and flow models and using that understanding to build better methods. I have worked on diffusion models for inverse problems and, more recently, on learning prediction targets for flow models.

At Caltech, I have worked with Prof. Anima Anandkumar and Prof. Bahareh Tolooshams. In summer 2026, I was a research intern in the Guibas Lab at Stanford. I am applying to PhD programs for Fall 2027.

News

  1. Nov 2026
    AnisoFlow is available on arXiv.
  2. Aug 2026
    EquiReg was accepted to TMLR.
  3. Jun 2026
    Received the Henry Ford II Scholar Award for the best academic record in computer science.
  4. Mar 2026
  5. Jan 2025
    DiffStateGrad was accepted to ICLR 2025.

Research

A grid of AnisoFlow text-to-image samples generated in pixel space, with legible newspaper, cinema marquee, and scoreboard text.

Learning Anisotropic Prediction Targets for Flow Models

Rayhan Zirvi*, Hansheng Chen*, Gordon Wetzstein, Leonidas Guibas

* Equal contribution.

arXiv preprint, 2026

A fully learnable prediction target for flow models that learns how much noise to keep in the target along different directions. Improves JiT from 1.90 to 1.55 FID on ImageNet 256×256. Largely removes patch artifacts in pixel-space text-to-image models under low-precision inference.

Paper Project Page

Schematic of the EquiReg principle: applying a transformation to on-manifold samples gives low equivariance loss, while off-manifold samples give high equivariance loss.

EquiReg: Equivariance Regularized Diffusion for Inverse Problems

Bahareh Tolooshams*, Aditi Chandrashekar*, Rayhan Zirvi*, Abbas Mammadov, Jiachen Yao, Chuwei Wang, Anima Anandkumar

* Equal contribution.

TMLR, 2026

A plug-in equivariance regularizer for diffusion-based posterior sampling that discourages trajectories from drifting off the data manifold. Improves reconstructions across linear and nonlinear imaging and PDE inverse problems.

Paper Code

Schematic comparing diffusion trajectories for inverse problems with and without DiffStateGrad: projecting the measurement gradient keeps the process closer to the data manifold.

Diffusion State-Guided Projected Gradient for Inverse Problems

Rayhan Zirvi*, Bahareh Tolooshams*, Anima Anandkumar

* Equal contribution.

ICLR 2025

A plug-in method for diffusion-based inverse problem solvers that projects measurement gradients onto a low-rank subspace of the current diffusion state. Filters unreliable gradient directions during sampling, improving reconstruction quality and robustness to guidance hyperparameters and noise.

Paper Code

Honors & Awards

  1. 2026
    Barry Goldwater Scholarship
  2. 2026
    Henry Ford II Scholar Award Caltech

Competitive Programming

  1. ICPC North America Championship (Honors) Southern California Regional Champion
  2. Codeforces Master top 40 in the US
  3. Meta Hacker Cup Round 3 Qualifier (Semifinalist) top 1% internationally