GenPeach AI

Nov 5, 2026
Last year, we started GenPeach, an AI research lab, because we believed there was a fundamentally more efficient path to building SOTA visual foundation models.
‍
Today, we're sharing the first results of what we've been building.
‍
In a blind study conducted by Contra Labs, 53 professional creatives evaluated our first model against 8 frontier image-generation models (such as GPT-Image-2.5, Meta Muse and others) across 100 human-centered commercial photography briefs. GenPeach won 71% of head-to-head preference votes. Evaluators considered 49% of its images ready for client delivery without further edits, more than twice the average of the competing models.
‍
Our model was trained entirely from scratch by a small team of exceptional researchers and engineers, using a fraction of the resources typically required for frontier model development.
‍
The result is an early demonstration of the research approach we set out to pursue.
‍
Rethinking How Visual Foundation Models Are Built
Scaling has been central to progress in generative AI. Larger datasets, more parameters, and greater compute have pushed model capabilities forward. But scale is not the only path to better performance.

At GenPeach, we're pursuing a different approach: rethinking model architecture, data utilization, and training methodology to build vertical visual foundation models with frontier-level performance. Our starting point is deliberately focused. Human-centered imagery is one of the most technically demanding domains in visual generation, with exceptionally high standards for quality and realism in professional production.

We believe that verticalization, combined with fundamentally different approaches to data and training, can deliver frontier-level performance with a fraction of the resources traditionally required. Our first model provides an initial validation of that thesis. The broader objective is to make this approach systematic and extend it to other increasingly demanding visual-generation problems.

A Small Team with Deep Frontier Research Expertise
GenPeach is built around a simple principle: a small team of exceptional researchers and engineers, with the technical depth and execution speed to take frontier research from first principles to production.

Our research is led by our CEO and Chief Scientist, who spent four years at Meta GenAI as a Staff Research Scientist, pushing the frontier of image and video generation. He co-led and led the development of Movie Gen, Meta's largest video generation model, and Imagine Flash, a real-time image generation model. His expertise spans the full lifecycle of frontier image and video models, from architecture design and large-scale training to inference optimization.

We've deliberately assembled a team that combines exceptional research talent, deep engineering expertise, and strong product sense. Research velocity matters. So does the ability to translate model improvements into capabilities professionals can actually use and commercialize. We believe the next generation of leading AI labs will need to excel at both.

Beyond Our First Model
Our ambition extends beyond a single model or category of imagery. We're building toward a broader set of production-grade visual foundation models, where exceptional quality, controllability, and efficient deployment are fundamental to the technology.

Our first model is an early demonstration of what a different approach to foundation-model development can achieve. We're building GenPeach to advance the frontier repeatedly, tackle increasingly complex problems in visual AI, and translate our research into commercially viable technology from day one.

Our first model is currently available to a small group of early users. We're gradually expanding access to creative professionals and teams.
Join the waitlist to be among the first to try it.

Read the independent Contra Labs evaluation
‍‍
Get Access  Apply  Contact





‍