About the group
Who are we?
We are a non-profit, volunteer-run AI research lab focused on rigorous, method-driven work in image diffusion, scientific machine learning, and lightweight model design. We produce peer-reviewed research and open-source models by combining experienced researchers, sponsored compute with demanding and novel AI research.
How we work
Bradbury Group operates as a distributed collective. Without commercial objectives or institutional overhead, every resource goes directly into compute and experimentation. Partnerships with sponsors such as Lambda Labs, Modal, and Weights & Biases give each project access to infrastructure that typically requires industry-scale budgets.
Project-Based Teams
Research is organized around discrete projects, each targeting a specific publication. Every project assembles its own team: machine-learning researchers with relevant expertise, domain specialists for scientific applications, and members of our standing core group who provide continuity across research threads. This ensures each project has a team built specifically for the problem at hand.
Parallel Execution
A typical project involves five to ten researchers working simultaneously - running ablations, implementing baselines, iterating on architectures, and preparing manuscripts in parallel. Coordinated execution lets us move from concept to submission significantly faster than traditional academic timelines allow, without compromising rigor or reproducibility.
Cross-Project Collaboration
Members contribute to multiple projects over time, accumulating co-authorship across active research threads as they contribute where their expertise fits. As workloads rise and fall, researchers contribute to one project, advise on another, or build shared infrastructure that supports the entire lab.
Research Direction
Project ideas typically originate from group leadership or senior researchers, often beginning with initial proof-of-concept work or early experiments. From there, teams are assembled to drive ideas through full-scale development, evaluation, and manuscript preparation. As projects mature, new research directions naturally emerge, leading to follow-on papers and additional collaborations.
Compute and Funding
Compute is arranged on a per-project basis through sponsorship agreements, research grants, and internally maintained resources -including group-funded specialized servers. Because no one draws a salary, the entirety of any funding is directed into compute and experiments. This efficiency enables project scales that would require substantially larger budgets in a traditional academic lab.
The Team
We maintain an active core of ~30 researchers at any given time, with a broader network whose availability fluctuates alongside academic calendars, exams, and industry commitments. This rolling structure allows the lab to scale project capacity dynamically and supports flexible contribution as members’ schedules change.
Our Principles
Authorship follows contribution. We track work transparently and credit based on actual output, not seniority or tenure. We select for researchers who can operate independently - designing experiments, debugging failures, and pushing projects toward completion. Expectations are high and the goal is always a finished, published result.