Something interesting is happening in the world of generative media. Not in the flashy, this-charactergot-its-fingers-right kind of way, but quietly — under the hood, in product workflows, and in infrastructure diagrams. The recent conversations and projects we’ve been part of have made something increasingly clear: we’re entering a second phase of generative AI, where speed, orchestration, and reliability matter more than paper benchmark scores.
Image generation is already there. Between models like Imagen 3, Flux, SDXL, and their fine-tuned cousins, there’s a growing sense of stability. Marginal improvements in resolution, fidelity, and stylistic control still matter, especially to commercial users, but the goalposts feel relatively fixed. Developers can now choose their model like an artist picks brushes: different tools for different jobs — from stylized avatars to pixel-perfect logos to archetypal product photos.
Video generation is a different story. It still feels like a knife fight in a phone booth, with new entrants leapfrogging each other every couple of weeks. There’s no single reigning champion. Sora, Luma, Runway, and others have all had their moment. That volatility is exciting, but it poses a serious challenge for entrepreneurs and researchers who want to build “on top” of video, not just with it. It’s hard to commit to tooling when the underlying capabilities might shift by next Tuesday.
This constant flux makes infrastructure oddly more interesting — and strategic — than one might assume. It also highlights just how much ground we’re still covering, even in areas that seem mundane. One of our early efforts, for example, was shaving stable diffusion’s cold-start latency from 19 seconds to something plausible for user interaction. Kubernetes orchestration alone introduced a ~5 second delay that just wasn’t viable for production. At that time, we had 8 GPUs. Now we routinely orchestrate workloads across tens of thousands, layered across multiple clouds and file systems, with caching handcrafted down to RAM and NVMe levels.
Those infrastructure decisions aren’t side quests. They determine what kinds of user experiences are even possible — and how quickly someone can go from “I wonder” to “here’s the asset.” For real-time generation and interactive workflows in particular, being able to serve cold-start workloads in milliseconds rather than seconds is a deal-breaker.
Fine-tuning is another area that’s getting more operationally interesting. It turns out that image generation workflows favor fine-tuning far more than language models do — not just 10x or 100x, but more like 1,000x more frequent. There’s a clear reason: brand consistency, character control, and creative stylization still require a tight feedback loop. One developer might fine-tune a model for a retail client’s product catalog; another might build a pipeline for virtual try-ons. The need for this kind of granularity makes the system dynamic in a very literal sense — the underlying models are often being shaped en route to deployment.
As generative workflows become more complex — chains of models doing things like background removal, segmentation, inpainting, 3D estimation, and final rendering — it’s striking how much tooling ergonomics start to matter. A good visual orchestration tool becomes not just a convenience, but a kind of compiler for creativity. Developers are building “apps” from models now, and they want more than console calls — they want legible interfaces for chaining logic, asset management, and iterative trial.
At the same time, marketplaces are emerging not from marketing pushes, but from usage patterns. Developers increasingly want a central place where they can browse viable video models, embed versions that are stable and fast, fine-tune if needed, and move on. On the flip side, model creators want to launch models into developer ecosystems without babysitting infrastructure, inference cost spikes, or janky web demos. If a closed model appears in our UI one day, it’s typically because enough developers asked for it that someone upstream said yes.
This isn’t just about supply meeting demand — it’s about platform integrity. We’ve bet heavily on customer proximity: small teams talking directly to users, engineers joining Slack channels and helping debug prompts or chain optimizations. There’s something fulfilling — and effective — about building systems this way. Not perfectly scalable perhaps, but visionary enough to evolve along with the landscape.
All of this — the competitive sprint in video, the workflow escalation in image, the quietly biblical amount of GPU being spun up and spun down — points to something bigger. We don’t think generative video is a fad. It’s not even a question anymore. Like mobile video before it, it’s an inevitable output mode. If you can promise, say, a tailored ad campaign with real-time rendering of scenes incorporating seasonal elements, or emoji-induced variations for consumer engagement, someone will say yes — and someone else will figure out how to do it cheaper next week.
If we had to circle a year on the calendar: 2025. That’s when this tips. Not just in quality, but in integration and utility. From real-time ads to interactive virtual experiences, from AI-powered game cinematics to everyday generative recreation, we’re getting close. The ideas are out of the lab. The systems are in Slack. It’s less magic now than momentum.
Image generation is already there. Between models like Imagen 3, Flux, SDXL, and their fine-tuned cousins, there’s a growing sense of stability. Marginal improvements in resolution, fidelity, and stylistic control still matter, especially to commercial users, but the goalposts feel relatively fixed. Developers can now choose their model like an artist picks brushes: different tools for different jobs — from stylized avatars to pixel-perfect logos to archetypal product photos.
Video generation is a different story. It still feels like a knife fight in a phone booth, with new entrants leapfrogging each other every couple of weeks. There’s no single reigning champion. Sora, Luma, Runway, and others have all had their moment. That volatility is exciting, but it poses a serious challenge for entrepreneurs and researchers who want to build “on top” of video, not just with it. It’s hard to commit to tooling when the underlying capabilities might shift by next Tuesday.
This constant flux makes infrastructure oddly more interesting — and strategic — than one might assume. It also highlights just how much ground we’re still covering, even in areas that seem mundane. One of our early efforts, for example, was shaving stable diffusion’s cold-start latency from 19 seconds to something plausible for user interaction. Kubernetes orchestration alone introduced a ~5 second delay that just wasn’t viable for production. At that time, we had 8 GPUs. Now we routinely orchestrate workloads across tens of thousands, layered across multiple clouds and file systems, with caching handcrafted down to RAM and NVMe levels.
Those infrastructure decisions aren’t side quests. They determine what kinds of user experiences are even possible — and how quickly someone can go from “I wonder” to “here’s the asset.” For real-time generation and interactive workflows in particular, being able to serve cold-start workloads in milliseconds rather than seconds is a deal-breaker.
Fine-tuning is another area that’s getting more operationally interesting. It turns out that image generation workflows favor fine-tuning far more than language models do — not just 10x or 100x, but more like 1,000x more frequent. There’s a clear reason: brand consistency, character control, and creative stylization still require a tight feedback loop. One developer might fine-tune a model for a retail client’s product catalog; another might build a pipeline for virtual try-ons. The need for this kind of granularity makes the system dynamic in a very literal sense — the underlying models are often being shaped en route to deployment.
As generative workflows become more complex — chains of models doing things like background removal, segmentation, inpainting, 3D estimation, and final rendering — it’s striking how much tooling ergonomics start to matter. A good visual orchestration tool becomes not just a convenience, but a kind of compiler for creativity. Developers are building “apps” from models now, and they want more than console calls — they want legible interfaces for chaining logic, asset management, and iterative trial.
At the same time, marketplaces are emerging not from marketing pushes, but from usage patterns. Developers increasingly want a central place where they can browse viable video models, embed versions that are stable and fast, fine-tune if needed, and move on. On the flip side, model creators want to launch models into developer ecosystems without babysitting infrastructure, inference cost spikes, or janky web demos. If a closed model appears in our UI one day, it’s typically because enough developers asked for it that someone upstream said yes.
This isn’t just about supply meeting demand — it’s about platform integrity. We’ve bet heavily on customer proximity: small teams talking directly to users, engineers joining Slack channels and helping debug prompts or chain optimizations. There’s something fulfilling — and effective — about building systems this way. Not perfectly scalable perhaps, but visionary enough to evolve along with the landscape.
All of this — the competitive sprint in video, the workflow escalation in image, the quietly biblical amount of GPU being spun up and spun down — points to something bigger. We don’t think generative video is a fad. It’s not even a question anymore. Like mobile video before it, it’s an inevitable output mode. If you can promise, say, a tailored ad campaign with real-time rendering of scenes incorporating seasonal elements, or emoji-induced variations for consumer engagement, someone will say yes — and someone else will figure out how to do it cheaper next week.
If we had to circle a year on the calendar: 2025. That’s when this tips. Not just in quality, but in integration and utility. From real-time ads to interactive virtual experiences, from AI-powered game cinematics to everyday generative recreation, we’re getting close. The ideas are out of the lab. The systems are in Slack. It’s less magic now than momentum.