How the House of David team made the show they were told was impossible

House of David tells the story of King David at the scale the Bible describes. And the Bible does not describe a modest story to bring into a show. Armies of 100,000 on each side of the Valley of Elah. Ancient cities, fortresses, desert landscapes, fire. Season 2 opens with an entire episode inside that battle, starting at dusk, moving through darkness, ending at sunrise.

The economics said no. "We had this core problem that we were told our vision was impossible," Erwin says. "That same conversation is happening all over our industry."

A mid-range streaming budget doesn't buy massive set builds, thousands of extras, or months of location work. It doesn't buy the dangerous shots either: horses charging directly at the lens, in low light and dust, on uneven terrains.

The choice was between shrinking the vision to fit the budget, or changing how the show got made.

Wonder Project's answer was not to swap filmmaking for AI. The rule on set stayed simple: anything that can be done practically, they do practically. Live action was shot in Greece with real actors, real department heads, and fewer than 200 extras at any given time.

What changed was everything running alongside the cameras. From the start of Season 2, the team planned generative AI into the pipeline as what Erwin calls "new principal photography", a digital second unit generating in parallel with the shoot. While the crew filmed, the team generated the wide shots the budget couldn't afford, extended locations, and amplified scenes in real time. Batches of generated material went to editorial and were reviewed like dailies; only the shots that made the cut were finished to broadcast quality.

The discipline behind it was unglamorous. The team trained on the tools daily, the way stunt performers drill. They kept a strict chain of title — no text-to-image, no text-to-video; every generation starts from assets the production owns and can prove it owns, tracked from source to final pixel. And they never treated AI as a finishing trick: generated elements were fused with live action and traditional VFX — Unreal Engine builds, Nuke, After Effects — so each technique covered the other's weaknesses.

It changed who got to collaborate, too. Decisions that usually wait for a post house months after wrap happened on set, with the department heads still in the room. Animated previs of entire sequences ran in production meetings so directors, stunt coordinators, and cast could see the scene before shooting it.

The result: a team of 600 doing work that would traditionally require thousands.

For the AI VFX team, the practical problem was volume and speed: dozens of environments, previs sequences, and background plates needed day to day, each requiring tests across several image and video models before one held up. Magnific became the platform the team used the most, one place to run generations on multiple models at once and keep everything in a central location, instead of scattering work across separate accounts.

The majority of the AI VFX team's work ran through it: immersive backgrounds for the volume walls, extending the canvas for establishing shots, and previsualization with motion. Directors used the platform live during studio shoots to improve and ideate on the volume wall backgrounds between setups. For example, one Season 2 sequence was shot outside and in the studio over four days, with much of the background AI-generated video, including green-screen elements, built on the platform with image, video, and editing models.

Magnific was the finishing layer and the discovery of the season. Upscaling had been in the stack since Season 1; Erwin has been using it for a couple of years. "I think it's art in up-res," he says. But on Season 2 the team found that Magnific's style transfer could take a frame from the show and transfer its look onto generated footage seamlessly.

That matching is what makes the hybrid invisible: the same sequence cuts between live action, traditional VFX, and generated shots without the seams showing.

For the studio, the other half mattered just as much as the pictures. "On season one and season two of House of David, Magnific's enterprise feature set really allowed the studios to be comfortable with the generation of this content — using these tools in a safe and ethical and effective way," Erwin says.

Season 2 premiered October 5, 2025, launching Wonder Project's subscription channel on Prime Video, which passed half a million subscribers in its first three weeks, ahead of the company's own goals. The season rolled out to all Prime Video subscribers worldwide on March 27, 2026.

The numbers behind the season: 253 AI-assisted shots, up from 73 in Season 1; more than 100 environments generated for the LED volume; builds that traditionally take 10–12 weeks and cost $15,000 to $200,000, compressed to about a week. For every VFX shot in the show, the team generated twenty times the material, giving editorial options a traditional pipeline never could.

At AWS's Culver Cup showcase in October 2025, Erwin screened the battle episode for an audience of filmmakers and challenged them to spot the generated shots. Horses charging past the lens: AI, built from show assets, too dangerous to film. Wide shots: practical extras blended with generated crowds. Most of the room couldn't tell.

"I'm proud of the level of scope and scale we put on the screen. I do not feel that we compromised," Erwin says. His script for House of David spans three seasons, and Wonder Project's slate is built on the same hybrid pipeline.