Robin Rombach

Overview

Robin Rombach is a machine-learning researcher and entrepreneur who works on diffusion-based image generation. He is the first author of the paper that introduced latent diffusion, the approach behind Stable Diffusion, and he later co-founded Black Forest Labs, the company behind the FLUX model family, where he serves as CEO.^[1][4]

The latent diffusion paper

In December 2021 Rombach and four co-authors posted "High-Resolution Image Synthesis with Latent Diffusion Models" to arXiv. The submission is dated 20 December 2021, and the author list runs Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser and Bjoern Ommer, with Rombach as first author.^[6] The paper was accepted at CVPR 2022.^[6]

The idea it set out is easiest to describe by contrast. A standard diffusion model learns to turn noise back into an image by denoising it step by step, but doing that directly on pixels is expensive, because a high-resolution image contains a very large number of values to predict at every step. Latent diffusion first compresses the image into a smaller learned representation, a latent code, runs the denoising process there, and then decodes the result back into pixels. The heavy iterative work happens in a much smaller space, which makes training and sampling cheaper while still allowing high-resolution synthesis. That is the method behind Stable Diffusion.^[1] The trade-off is that fine detail is bounded by how much of the image the compressed representation keeps.

At Stability AI

Rombach is the last author of the Stable Diffusion 3 technical paper, "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis", published in March 2024 and written at Stability AI.^[2] In machine-learning papers the final author usually signals a supervising or senior role. The subject — scaling a rectified-flow transformer backbone for high-resolution synthesis — marks where the line of work went after the original latent diffusion model.

Black Forest Labs

In 2024 Rombach co-founded Black Forest Labs with Andreas Blattmann and Patrick Esser.^[3][5] The company announced its launch on 1 August 2024 together with the FLUX.1 model suite and a closed seed round of 31 million dollars led by Andreessen Horowitz.^[7] Rombach serves as CEO.^[4]

At the launch of FLUX 3 in July 2026 he said, "You can't cheat reality."^[4]

Why the work matters

Latent diffusion is a design choice rather than a single product, but it changed what was practical. By moving the most expensive part of image generation into a compressed space, it made diffusion models for high-resolution images trainable and runnable at a scale that a pixel-space model of comparable quality would have struggled to reach.^[1]

The path from that paper to Stable Diffusion, then to Stable Diffusion 3, and then to a company building its own model family traces a common route in recent AI research: a method published at a conference becomes a widely used model, and the people who built it go on to start a company around it.^[1][2][7]