This is a car exploding on a bare studio stage — generated end to end from a single first frame and a single last frame. No client, no project: this is TDM R&D. Scroll to scrub it, then read how we got here.
TDM R&D capability test: a first-frame / last-frame destruction beat, played back frame by frame as you scroll. No client, no project. 97 frames from a single generated take.
On a night in Mexico in 2015, the crew of a James Bond film set off the largest stunt explosion ever recorded on camera: 8,418 litres of kerosene and 33 kilograms of explosives, close to three hundred detonations firing in succession over seven and a half seconds. When it was over, the crew celebrated. Not because it looked spectacular, though it did, but because it had worked the first time. There was no second take. There was never going to be one.
That’s the problem this article is about.
The shot you get one chance at
A practical explosion concentrates three production pressures into a single moment. It’s expensive, it’s dangerous, and you capture it once or not at all.
One take, no rehearsal
The Bond blast was overseen by Chris Corbould, the special effects supervisor whose credits run from the Bond films to Christopher Nolan’s work. He describes explosions as the most stressful element of his job, the part he only relaxes about once they’re finished. That’s not showmanship. When a set piece can only happen once, every decision made in the weeks before it has to be right on the night, because the location, the light, and the build are all consumed in the same seven seconds.
Safety is the line that does not move
Corbould is exact about priorities. “The safety of the stunt driver and stunt people is totally non-negotiable,” he has said. A practical destruction shot brings licensed pyrotechnicians, a fire-safety crew, permits, an insurance line, and a location you may not get back. None of that is padding. It’s the cost of doing the thing safely, once.
Why this matters to a budget
For a producer, the pressure isn’t abstract. It’s a scene the script wants, priced at a number the schedule can’t absorb, on a location the permit office may not clear, with a single chance to get it clean. That’s the real constraint, and it’s worth stating plainly before we get to what’s changed.
The Spectre (2015) compound explosion, a single-take practical effect and the largest stunt explosion recorded on film.
How this gets made without AI
Until recently a producer had two honest routes to a destruction beat, plus an older third. None of them was a mistake. Each was, and still is, a rational answer to the problem.
Practical pyrotechnics
The route above. You stage the real thing, with real fire crew and real permits, and you capture it in camera. It gives you a truth no simulation matches, and it carries the cost, danger, and one-take exposure that come with it. A German special effects firm puts the constraint in its own words: the effects often happen close to people, and because production costs are high, there is often only one attempt.
Full-CG destruction simulation
The digital route. A visual effects house builds the collapse as a physics simulation, rigid-body fracture, pyro, particles, run in software such as Houdini. This is deterministic and fully art-directable. An artist can dial the timing of a single fragment, the density of the smoke, the direction of the blast, and re-run it until it’s exactly right. What it costs is weeks of skilled artist time and render, which is its own kind of budget line.
Miniatures
The oldest path, and still alive: build the structure at scale and destroy the model. Slower and more specialised now, but real, and for some shots still the cleanest answer.
The point worth holding onto is that AI doesn’t correct an error here. These methods were rational. AI expands the option set. It doesn’t retire the craft.
Thirty minutes is usually enough to understand whether it makes sense for your project and what it would realistically take to implement.
Book a conversation →What has actually changed
Two things changed at once, and it helps to separate them.
A destruction shot has already shipped
This isn’t a promise about the future. In 2025, Netflix disclosed that a Buenos Aires building-collapse shot in the series El Eternauta was made with generative AI, produced with its own Eyeline Studios, and put on screen precisely because the shot was outside the show’s budget. It was reported as the first official use of generative video in final footage in a Netflix original. A destruction beat that would not otherwise have existed, made because the practical and CG routes were both out of reach, and delivered to air.
What the models can do today
On current video models, a workflow that defines the first frame and the last frame of a shot now produces a convincing main explosion beat, fast, and at a resolution high enough to composite on. Action beats in particular benefit from a defined end frame: it gives the model direction to generate dramatic motion rather than timid movement. Resolution has kept climbing too: some models now generate natively at 4K, and even where a beat starts lower it arrives as a strong plate to build on rather than something to rescue with an upscale later. That holds up both on the open web and in our own R&D — it’s the shot you scrolled through at the top of this page.
What they cannot do unaided
Here’s the honest limit. The main explosion works well as a starting point. The fine detail does not come finished: debris, secondary fragments, dust, embers, and other small particles are where the generated beat falls short, and they are best handled with additional compositing passes on top. That finishing method, layering smoke, debris, and ember passes over a base, is standard visual effects practice. The only new part is that the base beat is generative.
Cheaper explosions are not the point. The point is a destruction beat you were never going to be allowed to shoot.
AI is not a Houdini replacement
It would be easy to read the above as “Houdini, but cheaper.” It is not. A destruction simulation gives deterministic, frame-by-frame artistic control over every element of the effect. Current AI destruction is a different workflow, with different strengths and different limits, and it doesn’t offer that same control. Presented plainly, it widens what a production can reach. It doesn’t replace the tool that lets an artist art-direct every fragment.
Which shots are candidates
The useful question for a producer is not “can AI do explosions.” It’s “which of my explosions.” The line is reasonably clear.
Good candidates
A contained destruction beat with controlled framing: a single detonation, a partial collapse, a blast seen from a fixed and considered angle, with no close performer interaction. These are where the current workflow is strongest today, and where a generated beat plus compositing can carry the shot.
Poorer candidates
Sustained, performer-interactive pyro with foreground actors, long destruction events, or highly art-directed physics where every fragment has to land on a mark. These still benefit from traditional workflows, practical or full CG, and a producer is better served by hearing that plainly than by discovering it in post.
One direction worth watching
There is a hybrid approach still in exploration, not yet a shipped service: build a simple 3D animation or proxy simulation, use it as motion guidance for the AI generation, then finish with composited layers. The promise is meaningfully more artistic control without a full destruction-simulation setup. It’s early, and worth naming as a direction rather than a deliverable.
A note on compliance, kept proportionate. A pure environment or object destruction beat, with no generated people in it, does not trigger the performer-likeness rules in the German KI-Tarifvertrag. Where the pipeline is human-steered, shot design, frame definition, and compositing, it also sits on the right side of the Schöpfungshöhe question rather than being a bare machine output. Which model clears a given broadcaster’s provenance expectations is a project-by-project call, and part of the conversation before the work starts.
The one you can’t shoot twice
Corbould’s crew celebrated because the shot worked on the only take they had. That pressure hasn’t gone anywhere, and for many shots it shouldn’t: some things are still best done for real, once, with a fire crew standing by. What has changed is the set of beats that were never on the table at all, the collapse the budget would not cover, the detonation the permit office would not clear, the location you could not risk. For a contained, well-framed version of those, there is now a route that ends in a shot rather than a line item you cut. Knowing which of your shots that describes is the thirty-minute conversation.
