Andreas Blattmann

Overview

Andreas Blattmann works on generative image models. He is listed as the second author of "High-Resolution Image Synthesis with Latent Diffusion Models", the paper that is the basis of Stable Diffusion, and he is a co-author of the Stable Diffusion 3 paper, "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis".^[1][5][2] In 2024 he became one of the founders of Black Forest Labs, the company that released the FLUX.1 image models.^[3][4]

Diffusion in latent space

The latent diffusion paper was submitted to arXiv on 20 December 2021 and accepted at CVPR 2022.^[5] Its subject is the family of diffusion models, which generate an image by starting from noise and removing it step by step, with a network trained to predict what should be removed at each step. The problem the paper addressed is that doing this directly on pixels is expensive: high-resolution images contain a very large number of values, and every denoising step has to work on all of them.

The paper's answer was to move the diffusion process out of pixel space. Instead of training the diffusion model on raw images, it trains the model in the latent space of pretrained autoencoders: an autoencoder first compresses an image into a smaller, denser representation, the diffusion model is trained and sampled inside that representation, and the result is decoded back into an image. Working in this compressed space makes training and sampling cheaper while still producing high-resolution images.^[5]

The second ingredient is control over what gets generated. The paper adds cross-attention layers that let the model condition on text or on bounding boxes, so that the same generative model can be driven by a written description or by a layout of object positions rather than producing an unconditioned image.^[5] This is the mechanism that turns a diffusion model into a text-to-image model.

Stable Diffusion 3

In March 2024 Blattmann was a co-author of "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis", the paper associated with Stable Diffusion 3.^[2] The title points to two departures from the earlier recipe: rectified flow as the training approach, and a transformer backbone, both applied at the scale needed for high-resolution synthesis. Rectified flow, in general terms, is a way of learning the path that carries a noise sample towards a data sample, aiming for a path that is as close to straight as possible so that the transformation is easier to approximate and to integrate. The paper's subject is how such an approach behaves when the model and the data are scaled up.

Black Forest Labs and FLUX.1

Blattmann co-founded Black Forest Labs in 2024 together with Robin Rombach and Patrick Esser.^[3] The company launched its FLUX.1 models on 1 August 2024.^[4]

In the launch announcement of that day, the founders described the team's innovations as including VQGAN, Latent Diffusion, the Stable Diffusion models, Rectified Flow Transformers and Adversarial Diffusion Distillation.^[6] The list reads as a short history of the line of work Blattmann has been part of: a compression model for images, the move of diffusion into the compressed space, the open image models built on that method, a newer training formulation, and a distillation technique for making sampling faster.

Why it matters

The importance of this work lies in making high-resolution image generation from text practical to train and to run. The latent-space formulation is the basis of Stable Diffusion, one of the models that put open text-to-image generation in front of a wide public.^[1] The later papers and the company continue the same direction: a training objective aimed at more efficient transport from noise to image, and a family of released models built on it.^[2][6] Blattmann's position across these projects — second author on the latent diffusion paper, co-author on the Stable Diffusion 3 paper, and co-founder of the company behind FLUX.1 — places him among the people who carried the method from a 2021 research paper to systems that are released and used.^[5][2][3]