Prodia - The Fastest Image-Gen API in the World ⚡️
Plus: Co-Founders Mikhail and Monty on why edge compute has been a huge unlock...
Published 18 Apr 2025CV Deep Dive
Today, we’re talking with Mikhail Avady and Monty Anderson, Co-Founders of Prodia.
Prodia is a rapidly-growing image inference startup founded by Mikhail and Monty in 2022. Originally built as an internal tool for audio-related projects, the team pivoted to image generation after seeing massive demand and rapid adoption of their standalone image feature. Prodia’s core value proposition is simple: to be the fastest image generation API in the world - with an API that’s easy to implement and doesn’t require deep technical knowledge.
Today, Prodia powers hyper-fast image generation for customers ranging from well-known design platforms to consumer apps and enterprise platforms. Prodia’s infrastructure is optimized end-to-end—from authentication to model execution—to deliver real-time performance with minimal latency and high-quality output. With just a 4-person team, they’ve built a system that has placed them ahead of much larger teams on the vectors of cost, speed, and quality in image generation - including on recent third-party benchmarks from Artificial Analysis.
In this conversation, Mikhail and Monty talk about how Prodia was founded, why they’ve prioritized shipping product over fundamental research, and how they’re preparing for a future where the images are generated on the fly instead of from CDNs around the world.
Let’s dive in ⚡️
Read time: 8 mins
Our Chat with Mikhail & Monty 💬
Mikhail and Monty, welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on you and Prodia. What led you to co-found Prodia?
Hey there! We’re Mikhail and Monty, co-founders of Prodia. We originally started Prodia because we just couldn’t sit on the sidelines of the AI market. We were among the first 300 users of GPT-3, before ChatGPT and all of that, and we were blown away by what it could do. At the time, it was getting slower adoption than today, and we just couldn’t wait anymore—we jumped into it.
Originally, we made a bet on audio models because they were really bad at the time. We thought if we built a platform around them, we could catch the wave—as the models got better, our product would get better. We launched images to go along with our audio AI product for cover art, and that just took off. It went from 200 to 100,000 users in a month and a half. So we ditched all the audio stuff, went all in on images, and really saw the opportunity in creating an inference solution. It was really hard for us to get the right type of compute, optimize it, and build it into a product. Originally, that was just supposed to be a portion of the product.
We totally built the image feature as just a side thing to generate cover art and stuff like that for the audio. The discovery there was: it is really hard to find the right AI models, optimize them, and get them into your app—while still focusing on your product and user experience. We saw that as a huge need and went all in. The name Prodia stood for “Professional Media.” We wanted to make it easy to implement media into your app. So we built a super easy-to-use API. You don’t have to worry about GPUs, you don’t have to mess around with all the functions—you just focus on your user experience, not fiddling around with AI models.
How would you describe Prodia to the uninitiated developer or AI team?
We're the fastest and easiest inference solution for media in the world.
Who are your key users today? Who’s finding the most value in what you're building at Prodia?
We really serve companies where AI is just a portion of what they do—used to enable other core use cases. Our customers include Canva’s largest competitor, stock photo platforms, and apps where AI image or video generation is additive to the overall product experience.
Think of it like this: on the smaller end, we work with an app where users record their dreams. Previously, dream tracking was entirely text-based. Now, they can instantly generate a visual of what they dreamed about last night. The image isn’t the whole product—it’s just one part of the user experience. Those are our best customers. While we do support pure image generation use cases, we see the most traction when AI media is just one piece of a broader workflow or job to be done.
Which existing use-case for Prodia has worked best in your mind? Any user stories you’d like to highlight?
Our customer base is often B2B2C, sometimes B2B2B. We usually measure the impact we have on our customer’s product—not all the way down the chain. What we focus on is: are we helping you improve your product?
We’ll give you one use case. When we cut the speed of a key image generation flow in half, one of our customers saw an 80% increase in user engagement. Users were generating more, staying more engaged—and that became really powerful. That was a big win in that world.
Walk us through Prodia’s product. What use-case should new customers experiment with first, and how easy is it for them to get started?
Even on the mid-market enterprise side—we track this— our setup record is 7 minutes. We really want to focus on how to build the best API for someone who doesn’t have to know a whole lot about AI. A lot of people focus on selling models, but we’re really trying to sell a good experience for adding the media you want to generate into your app.
We mentioned one of our most recent customers integrated in 7 minutes - we’re really proud of that. We try to take away the pain so you don’t have to know about a trillion different models, how they compare, or how to do a bunch of tricks to get good output. We think we can help you get a really good output really quickly.
You’ll notice with Prodia—unlike a lot of other providers—we limit how many model options you have on purpose. Because you have a job to be done, and it’s usually not “serve every model that’s ever existed.” You want to generate an image for a particular purpose, and you want the highest quality while also maintaining really fast speeds and low costs. How do you do that? We just remove all that complexity for you.
Speed, latency, quality—you have to hit all three. Usually, you have to make trade-offs. We hit all three, and that’s why we focus. We don’t go into the long tail. Anyone can sacrifice one of those to gain more of the other. We really try to innovate and get better at all three at once.
You recently placed first on a set of third-party image API benchmarks set by Artificial Analysis. Tell us about this exciting result!
We reached number one in fastest image generation in the world—you can now generate a 1024x1024 image in 270 milliseconds. That unlocks a whole new world that wasn’t possible before. It’s now almost cheaper to generate an image than it is to load it from storage. Today, every image on a website is pulled from a CDN, but we think that over the next 10 years, 5–10% of the web’s images will be generated in real time—served from generative CDNs or possibly even from CDNless infrastructure. We’re still figuring out what to call it.
The idea is edge compute, generating the image just for you, instantly, based on who you are—without needing to ship anything across the internet or rely on heavy caching or storage. The unlock is personalization. Every person visiting a site can have a different visual experience. We recently did a demo where, based on your IP address, the landmarks in an image would change. If you’re in Germany, it shows one thing. If you’re in New York, something else. That kind of personalization used to be too expensive. Now it’s not just possible—it’s fast and cheap.
Could you share a little bit about how Prodia actually works under the hood? What is is about your architecture or system that enabled you to achieve this result?
It’s not one thing—or even ten things—that made us the fastest. For years now, we’ve focused on making our entire system ultra low latency. And that’s not just a buzzword. When you hit our API, we don’t touch a single database, we don’t make a single extra network call—you go right through. We use cryptography instead of touching a database to authenticate you, because databases are slow, blocking, they get congested. It gets assigned directly to a compute node so it doesn’t queue. We spent a lot of time optimizing this scheduling to be maximally fast and efficient—low cost and instantaneous.
And then once we actually get to the inference, a whole load of stuff happens. We compress the model with quantization. We compile it in a special way that we worked with some NVIDIA engineers on, into a binary that eventually runs as PTX assembly. All machine learning is matrix multiplication. If you understand that, it just becomes about how quickly you can crunch numbers. A lot of people only focused on that portion—on the ML side. We spent a good bit of last year just making sure the data moves around the internet fast. We found that to be the most important bottleneck. To the point where now we’re optimizing in tens of milliseconds. In the past, you used to try to optimize for seconds or hundreds of milliseconds.
It’s also important to point out we specifically picked out the best partners on the compute side to enable that. And that ended up being Lambda Labs and Google. We found that a GPU is not as commoditized as it seems on paper. We’ve had to disqualify a lot of compute providers. Once you actually run it, there are so many variables—how their networking is set up, how well they segregate your hardware, the reliability of the infrastructure. There’s just so much under the hood that we had to test, disqualify, and go through this painful process to figure out how we can squeeze out as much performance as possible. It’s part of what people pay us for.
How is your team balancing the need for shipping great products vs. the temptation to focus on AI research? Has open-source helped with this?
At Prodia, we made a strategic decision not to invest in the research side and instead partner with the brightest folks in the world. We’re talking about groups like Black Forest Labs with the Flux models, Recraft, Nvidia—we’re working directly with their researchers. We’re not trying to compete on the model research front. Our focus is purely on making a product that’s usable. It’s one thing to write a paper, it’s another to get something into production. The real question is: how do you take a model or research paper and turn it into a production-grade experience? That’s where we shine, and that’s where all of our energy and resources go.
To add to that—we benefit massively from the open source community. There’s incredible work happening there, and we’re building on top of it. Something we believe deeply is that the power of the open source community is so vast, diverse, and fast-moving. If you can learn quickly and leverage that, you unlock a massive advantage. That’s the key with open source—it’s not just about code, it’s about velocity.
We’re really inspired by Zuckerberg’s concept of a “learning organization”, and we’ve internalized that at Prodia. We focus heavily on continuous learning because in the age of AI, speed of execution is everything. You can’t compete unless you’re constantly evolving and staying ahead. We’ve found the open source world has massive structural advantages. Why was Llama open source while OpenAI and Google kept things closed? Because Meta doesn’t sell cloud services—so their incentive is to rally the community to build tooling and infrastructure around their models.
You could say it worked - the majority of production image generation today isn’t happening on closed-source models. DALL·E 2 and 3 never really took off in production settings. Even with the recent Giblify trend—those are novelty use cases. In the real world, OpenAI isn’t being used in production for image generation because you can’t access the good models, and even if you could, it’s just too slow and expensive. It takes 20–30 seconds per image. That doesn’t work for real product use cases.
What has been the hardest technical challenge around building Prodia into the product it is today?
The hardest part is really having the breadth across all these verticals—from networking, to CUDA, to inference, to rearchitecting machine learning models. There’s just so much involved. The challenge is pushing each of those pieces just past the current frontier, all at once.
A lot of our edge actually comes from deep knowledge of what people would call “boring” stuff—not just the AI side. One of our engineers, for fun, reads protocol documentation. He knows how HTTP works better than 99.99% of people in the world. And we’ve used some really clever, unconventional ways to leverage that protocol that most people wouldn’t even think of.
On the infrastructure side, our head of Infra Stefan literally owns an IP block just for fun. The infra team understands how data moves at a fundamental level—how it flows through Ohio vs. Texas vs. LA, how the global interconnect fabric works. We’re probably one step below high-frequency trading in terms of how much innovation and precision we put into moving data across the Internet.
One cool example: our engineer Caleb came up with something based on his deep RFC knowledge—bidirectional multipart form. Almost no one does this. In fact, ChatGPT won’t even generate code for it because it doesn’t think it’s a real use case—it’s that rare. But it ends up being absolutely ideal for what we do.
How do you see Prodia evolving in the next 6-12 months, given these results? Any technical or product developments that your customers should be excited about?
So the next challenge—once you’ve nailed speed—is how to get that same speed and better customization, better character consistency. One of the tradeoffs of real-time generation today is that you give up some of the tools that enable deeper control, like LoRA or ControlNet. Those are what let you get much closer to exactly the image you want. But they’re hard to incorporate when you’re optimizing for real-time.
As we work those capabilities in, the real-time image experience will get faster and more customizable—higher quality, more consistent. That’s going to be a big focus for us. At the same time, speed will still be a huge priority. It’s a win-win across the board. We spend less on compute, so prices go down for our customers. The user experience improves. And for the end user, it just feels seamless—they’re not stuck waiting. So speed continues to drive value on all fronts.
Lastly, how would you describe the culture at Prodia? Are you hiring, and what makes the current team special?
This is going to be shocking—we’re the SEAL Team 6 of AI... but really SEAL Team 4, because there are only four of us. Two of us are here right now! And what’s wild is how competitive this space is. Everyone else has raised hundreds of millions and has teams of 20+ working just on optimization. Not even full product teams—just optimization! And we’re matching or beating them with four people.
Everyone on the team has a superpower. Caleb is one of the best in the world when it comes to APIs and protocols. Stefan is elite when it comes to infrastructure, hardware, and data center knowledge. Monty is an incredibly rare bridge between engineering, product, and business—someone who can go deep technically and still intuitively understand the user experience. And then I bring the voice of the customer into everything we do.
Anything else you'd like our readers to know about the work you’re doing at Prodia?
The rallying call is: if you're interested and you really geek out about performance and latency, talk to us. We don't even need to sell you anything. We love conversations—we love talking open source, we love talking about what we do. And if we can help you, even if you don't become a customer, we’d love to connect. Hopefully we both learn something.
Whatever it is, this is our passion, it’s our core, and we’re the best in the world at it right now. But to stay there, we have to keep learning. Like I mentioned earlier, being a learning organization is part of our DNA. That’s why this call to action is ultimately about learning.
Conclusion
Stay up to date on the latest with Prodia, learn more about them here.
Read our past few Deep Dives below:
If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email (newsletter@cerebralvalley.ai) or DM us on Twitter or LinkedIn.