---
title: "The man who made deep networks trainable now argues image generators forgot how to denoise"
date: 2026-11-16
category: Builders
site: NeuroAI
canonical: https://neuroai.site/a/na-people-kaiming-he-denoising
language: en
---

# The man who made deep networks trainable now argues image generators forgot how to denoise

> ResNet's lead author left Meta for MIT, then added a part-time DeepMind title. His new argument is unglamorous: image generators should go back to plain denoising.

In 2015, a plot in a computer-vision paper showed something that should not have been possible: a 56-layer network was doing worse on its own training data than a 20-layer one. It was not overfitting. The deeper model was simply harder to optimize, and adding layers made the result worse.

The fix turned out to be one line of arithmetic. Instead of asking a stack of layers to learn a mapping, you let it learn the difference from its input — add the input back at the end. That is a residual connection, and it is now inside almost everything: Transformers, AlphaFold, the diffusion models that make images, the video models that make clips.

The person who wrote it down, Kaiming He (何恺明), is now an associate professor with tenure at MIT's Department of Electrical Engineering and Computer Science, holding the Douglas Ross (1954) Career Development Professor of Software Technology chair. Since June 2025 he has also been a part-time Distinguished Scientist at Google DeepMind, a role he lists alongside the academic one rather than instead of it.

His current argument is smaller and stranger than ResNet, and it decides the cost of every image you generate.

## The fix was an addition

ResNet's contribution was not a new kind of neuron or a bigger dataset. It was a reparameterization: make it easy for a block to learn the identity function, and depth stops fighting you.

The paper, "Deep Residual Learning for Image Recognition," won the CVPR best paper award in 2016 and collected the conference's Longuet-Higgins Prize for test of time in 2026. On He's own homepage he notes that a *Nature* article described it as the most-cited paper of the twenty-first century, and that his publications passed 700,000 citations as of May 2025 — a number he reports himself, from Google Scholar.

The same pattern shows up across his record. Faster R-CNN, written with Shaoqing Ren, Jian Sun and Ross Girshick, took the NeurIPS test-of-time award in 2025. Mask R-CNN took the Marr Prize at ICCV 2017; the PReLU paper took ICCV's Helmholtz Prize in 2025. His first published paper, on removing haze from photographs with a dark-channel prior, won CVPR's best paper in 2009.

What those papers share: none needed new hardware. Each removed a specific obstacle other people were stuck behind.

## A career that moved in straight lines

The biography is unusually linear for a field that reshuffles every three years. Undergraduate degree at Tsinghua University (清华大学), 2007; PhD at the Chinese University of Hong Kong in 2011, supervised by Xiaoou Tang (汤晓鸥), who later founded SenseTime; Microsoft Research Asia (微软亚洲研究院) from 2011 to 2016; Facebook AI Research from 2016 to 2024; MIT EECS from 2024, with tenure following and the DeepMind appointment in 2025.

He never ran a company. When much of his cohort went to startups, he stayed in labs — and the work he produced there ended up in everyone else's products anyway.

## What a tenured professor actually ships

His group's output is easy to check: nearly every entry on his publication list carries arXiv, code, or a project page.

- **MAR (NeurIPS 2024, spotlight)** showed autoregressive image models do not need a discrete tokenizer; a small diffusion head can model continuous tokens directly, removing a stage from the pipeline.

- **MeanFlow (NeurIPS 2025, oral)** learns an *average* velocity field over a time interval rather than an instantaneous one, so a sample can be produced in a single network evaluation instead of tens of steps.

- **"Back to Basics: Let Denoising Generative Models Denoise" (CVPR 2026; tech report November 2025)** is the argumentative one, written with Tianhong Li.

He also teaches 6.7960, Deep Learning, in MIT's fall 2026 term.

## The denoising argument, in plain terms

Modern image and video generators are trained by a denoising objective: take a clean sample, add noise, ask the network to undo it. But most large systems do not ask the network to output the clean image. They ask it to output the noise, or the velocity that moves a sample along the noise schedule, because those targets behave better numerically at the start of training.

The "Back to Basics" paper argues that this detour has become expensive. If the network's job is to predict velocity, part of its capacity is spent on bookkeeping rather than on the thing you actually want. Ask it to predict the clean signal instead and quality improves, with the gap widening as models grow and data gets higher-dimensional — the direction video and 3D generation are heading.

The practical consequence is large. If you can hold quality with fewer parameters or fewer sampling steps, generation gets cheaper per image and per second of video — the difference between a demo and something people use daily.

MeanFlow pushes the same lever from the sampling side. Standard diffusion spends dozens of network evaluations per sample; if the model learns the average flow across a step, one evaluation can be enough.

## Why this matters outside the lab

**Anyone paying for inference.** Image and video generation is priced in GPU-seconds. A one-step sampler, or a smaller network at equal quality, changes the unit economics of every app that sells generation.

**Anyone building on open weights.** His group's choices tend to become defaults in other people's repositories within a year. The tokenizer-free recipe in MAR matters to anyone who wants language-model-style scaling on images without maintaining a separate vision tokenizer.

**Anyone watching where senior researchers go.** He trained in mainland China, took a doctorate in Hong Kong, and built a career across US industrial labs before landing a tenured chair — a path that exists but is narrow.

## Honest limitations

This profile is built from primary documents — his MIT EECS faculty page, his personal page and publication list, the arXiv records behind the papers — plus press coverage of the DeepMind appointment. I did not interview him, so his reasoning here is my reading of the papers, not his own words; there are no quotes for that reason.

The citation figure and the "most-cited paper of the century" claim are self-reported on his homepage, citing a *Nature* article I did not re-check. Award listings are his and MIT's; the 2025–2026 prizes are recent enough that third-party summaries lag.

The technical section simplifies a live argument: whether clean-signal prediction wins as the default for very large video models is unsettled, and I have no visibility into his DeepMind work. Popular Chinese-language biographies add details — a perfect gaokao score, a specific undergraduate major — that I could not confirm from institutional sources, so they are omitted.

## What readers can do now

- **Read the November 2025 tech report and its code** before accepting either side of the prediction-target debate; it is short, and the ablations are the argument.

- **If you fine-tune diffusion models**, run the cheap experiment: swap velocity prediction for clean-sample prediction at your model size and compare samples per second at equal quality.

- **Treat single-step samplers as a cost lever, not a curiosity.** When you price a generation feature, ask your vendor which sampler and how many steps; the answer often explains the quote.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-kaiming-he-denoising
Free to quote with attribution and a link to the original.
