AI Video Already Solved the Hard Part. Here's the One Thing It's Still Faking.

Every AI video company right now will tell you they've solved consistency. Scroll their landing pages and you'll see the same promise in different fonts: lock your character, keep your product steady, generate a whole campaign without anyone noticing it came from a model. Then you read their actual documentation, the help center guides written for people who use the product every day, and a different story shows up.

Build a reference sheet. Generate a hero image first, and don't animate it until you're happy with it. Write a character block and paste it into every prompt unchanged. Use the last frame of one clip as the first frame of the next so the cut doesn't show a seam. Compare every new shot against reference image one, never against yesterday's shot, because drift that's invisible clip to clip becomes obvious in a twelve-up grid.

Underneath the branding, that's just a production process, a good one even, that a human has to design, maintain, and check, shot by shot, forever.

The tell is in the instructions

I went looking at how the industry teaches people to solve this, and the pattern holds across tools: build a character as a still image first, freeze a written description of it, and never let the model improvise from memory. Treat every shot like it has amnesia, and compensate by handing it the same locked inputs every time. One continuity guide lists six layers you have to track by hand across every shot, from identity to cinematography. "If one layer changes without a story reason," it notes, "the viewer notices a break even when the individual shot looks attractive."

Read that list again. It's a job description, the kind a creative director and a script supervisor hand out on a real shoot. It's just been handed back to the customer, repackaged as a "workflow."

Generation is not the bottleneck anymore

The models are genuinely good now. Runway's own filmmaking guide admits what AI "can't do is tell the story. A film still needs a filmmaker." But for B-roll and mood pieces, generation quality has stopped being the thing standing between an idea and a finished clip. Need a lonely road at dusk? You'll have it in seconds.

The problem shows up the moment you need the same thing twice. PromptWise's review of marketing-grade AI video tools names the real complaint marketers have, and it isn't "the video looks bad." It's that "the product changes shape between shots, the mascot's face drifts, the colours wander." A thirty-second ad for a specific watch needs the same watch in every shot, and that level of product continuity is still where these models struggle.

Nobody buys a single gorgeous shot. They buy a campaign, and a campaign needs the product to look like itself in frame one and frame twelve. Generation quality got solved while nobody was looking. Continuity didn't.

The gap between a demo reel and a campaign

A demo reel only has to show the best clip out of fifty. A campaign has to hold together across every clip, every platform, every week, for as long as the brand keeps showing up in people's feeds. The industry has been selling solutions to the first problem as if they'd solved the second.

Marketers burned by a "just prompt it" pitch know this instinctively. They got one beautiful clip and three that looked like a different brand made them, and concluded AI video "isn't ready." The tools weren't lying about what the model can generate. They were quiet about what it takes to make that generation usable at scale: someone has to build the reference library, approve the still before it becomes a video, catch the drift before it ships, and decide what counts as "on brand," then hold that line over time.

That work didn't disappear. It just moved from "part of the product" to homework you do before you open the product.

Where the real work still has to happen

If generation is mostly solved and continuity is mostly manual, the obvious question is who automates the boring part. Not a flashier model, I'd guess. Whoever builds the infrastructure nobody wants to build because it isn't exciting: reference libraries that persist instead of living in someone's folder, approval built into the workflow instead of bolted on as a separate step, a style guide that works like code the system enforces instead of a paragraph someone pastes into a prompt.

Call it the continuity and brand-governance layer. Unglamorous, and as far as I can tell, the real bottleneck left in this industry. One of the clearer guides on this put it well: the fix is to "save the process as a workflow," so consistency "travels with the process instead of living in your memory." That's correct, and it's exactly the gap most AI video products leave for the customer to fill themselves.

Why we built Popcorn as a system, not a generator

This is the problem we started with at Popcorn. A prompt box, however good the model behind it, still leaves the hardest part of the job with the marketer: deciding what the brand looks like, holding that line across every piece of content, and catching it when something drifts before it goes out the door.

We built Popcorn to carry that whole chain, not one link in it. It learns the brand and the business, finds the stories worth telling, and turns existing content and IP into new concepts before anyone writes a script. Those concepts become creative strategy, then scripts, then video and motion graphics, cut together through a full timeline with the same shot-by-shot discipline a real production team would apply, built into the system rather than run by hand. Then it publishes, measures what performed, and feeds that back into what gets made next.

The part I care about most doesn't show up in a feature list: nothing goes out without the customer's permission. AI does the production work. The human stays in charge of what gets approved and sent back. We treat that as the whole premise, not a checkbox we bolted on. The industry's own workaround for consistency, approve the still, lock the reference, check it against the brand standard, is correct. We just don't think it should be a chore a marketer performs across five tabs. It should be built into the system, with a real approval step where a person looks at what's about to go live and says yes or no.

The tools that win the next phase of this industry won't be the ones with the most impressive clip in their demo reel. They'll be the ones that handle the tedious, essential work of keeping a brand looking like itself, with every new release, so the people running the brand can spend their attention on what's worth saying instead of whether the mascot's face held together between shots three and four.