The Impossible Project, Made Possible

How Mile 80 uses OpenArt to massively expand its output, and what becomes possible once the workflow changes

For Ryan Sheffer, the relationship with OpenArt started with one feature that did most of the heavy lifting. "The number one feature for the longest time we were using was start frame, end frame, make a video," he says: generate the first frame of a shot, generate the last frame, let the model fill in the motion between them. Simple as that sounds, it's how Sheffer's studio, Mile 80, has quietly put AI-generated shots into work for some of the most recognizable brands in the world. On a recent BMW show, Mile 80 delivered videos for Mini Cooper and Rolls Royce, and by Sheffer's account, most of the BMW-related output that season had at least one shot generated or partially generated with OpenArt, the kind of high-end clients, in his words, "you wouldn't think had AI in their videos."

Sheffer founded Mile 80 nearly a decade ago after years of working as what he calls "the nerdy guy" in post-production and visual effects. The studio built its name in live events: 360-degree screens, spatial storytelling, the kind of shows where a single project might need hundreds of assets rendered at 30K resolution. It's a business built on technical difficulty, and Sheffer has spent his career finding the next hard problem before everyone else does. That instinct is what brought him to AI early, and it's also what makes him refreshingly honest about where it actually helps.

Why start frame, end frame changed everything

Sheffer sat down with us to walk through what happened once his team stopped treating OpenArt as a standalone generator and started building around it.

The shift came when OpenArt opened up its MCP server. Mile 80 began layering its own interface on top, using Claude to manage the entire creative process, from building shot lists to critiquing individual frames, and routing the actual generation requests to OpenArt underneath. "We've started like putting UI on top of it," Sheffer explains, "basically we're only building now with Claude, which is very interesting. So our UI is Claude for OpenArt these days."

In practice, that means Mile 80's artists talk to Claude the way they'd talk to a coordinator on a physical set. They define characters, environments, and reference images, then Claude builds what Sheffer calls a production bible and shot list, and routes generation requests to OpenArt for the actual start and end frames. Two of his artists, Camila and Francez, built a custom Claude skill that takes a script, breaks it into individual shots, and fires off the OpenArt requests automatically. "It works really well," Sheffer says.

One of the pieces Sheffer is proudest of is a Rolls Royce animation Mile 80 built entirely with OpenArt-generated frames, the kind of luxury-brand work he says has no real precedent for AI involvement. "I can't think of a brand that is higher end than that," he says.

OpenArt's contribution here goes beyond just start and end frames. Rather than generating a finished background, Mile 80 prompts OpenArt to build visuals with a green screen already baked into the shot, then relies on its own compositing tools, the same ones it's always used, to pull that green screen out and layer the shot into the scene. It's the detail that lets Sheffer's team keep working the way a VFX shop always has, just with a generated element standing in for what used to be a live-action plate. The same approach produced a character Mile 80 built for LangChain, an animated figure named Newton generated with a green screen baked in and then composited out, set to appear as part of a New York City subway campaign.

The project that should've been impossible

The clearest example of what that pipeline makes possible is Gammarama, a live event series the client wanted positioned as something like the new TED talks: more serious than the brand's usual output, without losing its sense of fun. Sheffer's team had five weeks to develop the entire visual identity and then deliver on it. "There was only five weeks to develop the entire visual identity," he says, "what the deliverables would be, each of the video scripts, each of the videos, storyboards, and then execute and deliver for a large screen display."

Sheffer is direct about what that timeline would have meant without an AI-based pipeline. "It would have absolutely been impossible," he says. "There's no way that we could have done what we did." The team would have had to scale the concept down to something simpler and less distinctive just to hit the deadline: fewer visual styles blended together, fewer animated elements on screen, something closer to a straightforward edit than the layered, stylized identity the client got instead. He credits his lead animator, Fran, with pushing the team's AI workflows forward enough to make the ambitious version possible at all.

What made the timeline work wasn't a single clever trick so much as the team's willingness to treat "if this were possible, how would we do it" as the starting question, rather than "can we even attempt this." Green screen and alpha-channel generation handled the layering, while Claude, connected through MCP, handled direction and shot management. And because OpenArt could turn a concept into a usable start and end frame quickly, the team could iterate on a look before committing real production time to it.

The difference between craft and slop

Sheffer's read on generative AI is more particular than the average AI-optimist take, partly because he's been burned by the hype cycle before. Back in 2013, before founding Mile 80, he ran a company that automatically generated short news videos from trending social posts, pulling content based on hashtag and location spikes and stitching it into a reel with an early large language model. It worked, in the sense that it got views. It also produced, by his own account, a video about a viral photo of a slice of pizza at the World Series that did numbers, and that was the moment he started backing away from the business. "I did see AI slop immediately as the core issue," he says. "That's funnily enough kind of why Mile 80 was started."

That history is why he's specific about what AI tools are actually good for. "The generative AI tools are the world's fastest, and in some ways dumbest, employees you have," he says. "You have to be incredibly specific with what you ask for. And if you correctly ask them for it, they will give it to you." He compares the goal to how David Fincher uses visual effects in nearly every shot of every film he makes, so seamlessly that audiences never register them as effects at all. "No one thinks of him as a visual effects director ever," Sheffer says. "Why? It's so good. It's integrated into the scene." The alternative, in his view, is the same failure mode people associate with the least-liked Marvel visual effects work: technically present, but not actually serving the story. "That's the same thing as AI slop," he says. "It's not understanding how to use the tool."

He's equally clear about where he draws the line for his own use of it. "I don't ever want API access to OpenArt to generate a hundred thousand videos of a banana telling a joke," he says, and he means it as more than a joke: mass, ungoverned generation is precisely the failure mode he watched sink his last company.

More ambition, same deadline

Sheffer doesn't sugarcoat what faster generation actually buys Mile 80: not more clients, but far more ambitious work inside the deadlines they already have. "I found that it's faster," he says, "but the project that will now have Godzilla didn't previously have the time to make Godzilla the other way. Your capacity to go from idea to visualization is much improved." A concept that would once have gotten quietly simplified down to fit a deadline can now stay as ambitious as the brief actually called for.

The one constraint that hasn't moved is client review. "Even if we can do a video in a day, it still has to go up the chain," Sheffer says, and that approval cycle sets the real pace of a project no matter how fast the generation step gets. It's a distinction worth sitting with: the value here isn't a shorter timeline, it's a much higher ceiling on what fits inside the timeline you already have.

The next hard problem

The newest thread Mile 80 is pulling on is live action. Rather than generating a scene from scratch, the team is testing what happens when real footage, shot quickly and simply, gets AI effects, creature work, or set extension added in post. Sheffer thinks there's real room to run here: "I could see a workflow for sure of filming with an iPhone and getting out high quality Hollywood blockbuster style," he says, enough that he half-jokes about the approach eventually undercutting the traditional camera market.

The limiting factor, in his view, is still human emotion. "Our eyes are so used to, and our brains are so used to, knowing that's a human being, that it's always the hardest thing to fake," he says. That's part of why he's also been thinking through quicker face-replacement pipelines, layering real webcam performance onto a generated character rather than trying to generate the performance itself. It's a fitting place for the story to land: even the team most willing to push a pipeline as far as it can go still draws the line at the same place, the parts of the work that come from being human.