Six Weeks, $1,000 a Minute: Henry Daubrez on Making Junkyard King

Episode two, not three.

AI LIVE was back for its second live show in two days. The guest was Henry Daubrez, Google Labs' first resident filmmaker, there to premiere the next chapter of Junkyard King, his animated series made with invideo, who sponsored the session. One correction before the start: the first instalment, which Henry had numbered episode zero, is the pilot, so this is officially episode two.

Henry and the team go back a long way. Chris had met up with him again at Google I/O and at Upscale, and this was a reunion for Paul too. At the time of the Coca-Cola meet-up they both remember, Henry was still living in Belgium. He has since moved to Utah, and Chris noted how much of the AI work, the clients and the companies now come from the US, with Europe trying to keep up.

The weekly show had already featured Junkyard King more than once. Chapter 0 drew 14 straight 1s on episode 118, and the hosts usually played the first five or ten minutes of each episode before sending people to Henry's YouTube for the rest. This time the whole episode would play, uninterrupted.

Borrow the codes, write every word.

Junkyard King began as a test. Henry has been using AI for about five years, and he never does research and development for its own sake. He always keeps a project running behind the experiments, and this one started as a way to see what Seedance 2 could do when it came out.

“I don't like just doing R&D to do R&D. I like to have a project behind, so that I can build up something.”

Henry Daubrez▶ 5:13

The raw material is public domain: the legend of King Arthur, with its sword in the anvil, twisted into something more personal. The setting is the childhood Henry grew up in, somewhere between the 80s and the 90s, watching films like Explorers and The Goonies, where kids ride BMXs and get up to no good until dark. Details like Artie's yellow Walkman come straight from his own memories.

Chris asked why he built on something out of copyright rather than something original. Henry's answer is that it is original. He writes every episode himself and borrows only the codes people already know, the way Star Wars borrowed knighthood. With so much content around, familiar codes give an audience something to cling to while the story goes somewhere new. He compared it to what Kavan the Kid does with The Chronicles of Bone.

“I'm still writing every single episode to make sure that I go where I want to go. I think I'm just borrowing from the codes which are known from those IPs.”

Henry Daubrez▶ 7:46

Chapter 1 was made with invideo Agent One, and this episode moved to Agent Two partway through. Henry calls it an evolution, not a new workflow. The big gain is memory: he feeds in the script, and the agent keeps track of everything from episode zero onwards, so he can ask it about where the story has been. He also watches for the model's habit of agreeing with whatever you say, and sometimes closes the agent altogether to fine-tune by hand in the composition view.

“There's always this sycophancy with models. Most likely going to tell you that you're right, whatever you say. Try to avoid that.”

Henry Daubrez▶ 11:50

Like water, not fire.

▶ Watch Junkyard King, Chapter 2: The New Guy

The episode picks up two nights after Artie found a magic sword in the old switchyard. His best friend Kai saw the pink flashes, the talking mechanical crow and the woman in armour, and Artie has to decide what to do about him. A long confrontation in the school bathroom ends with Kai swearing on his walkie-talkie to keep quiet, and joining in.

At the yard, Socket explains how the doors between worlds work: they only open when the blade resonates with a world's own note, and Artie has been brute-forcing it. Kai offers to listen for the frequency. Socket's lesson is to go like water, not fire. Overdoing it brings a giant creature down on them, and Artie ends up opening a door in the middle of the fight and using it instead of fighting. The last beat, a voice on the walkie-talkie, sets up the next chapter.

“This is probably the best animation I've ever seen made with AI. I really mean that too, Henry.”

Chris Branch▶ 27:52

It answered Andrew's complaint the day before that none of the XPRIZE entries were animated. Andrew called it by far the best animation he has seen, and the hosts picked out the ET feel of the bike, the basket and the rolling camera, the classroom, the character designs and the voices.

One pass, seven minutes.

Anita's question came first: how do you keep the voices consistent? Part of it, Henry says, is something he can't fully explain. Once a character's voice lands and you keep working in the same environment, the model tends to keep delivering a similar voice, perhaps because of how it reads the character's design. One of the first voices he got reminded him of the late Star Trek actor Anton Yelchin, and he liked it.

Then he locks it with a clone. For every character he goes back to ElevenLabs and trains a voice clone. Usually he keeps the delivery the video model generated, because he likes its intonation, and runs it through the voice changer. On this episode he changed how: instead of converting line by line, he exported long chunks of the edit starting at zero seconds, voice-changed each one in a single pass, and dropped it back into the timeline already in sync.

“I would export the entire audio track at once, voice change the entire thing, like five, six, seven minutes straight, and bring it back in the editor and then just recut exactly where I know they were talking.”

Henry Daubrez▶ 31:45

Socket is different. His voice comes from the ElevenLabs library, with a pompous, Snape-like tone, and Henry adds the walkie-talkie crackle by hand after every line so it sounds like it is coming out of a small speaker. When the voice changer gives strange results, he falls back to text to speech and rebuilds the line by hand.

The bathroom was fine. The legs were not.

Asked which scene he was most nervous about, Henry started with a first. This is the first episode where he hired someone to help: Julie Wieland, credited at the end. They split the scenes, and Julie took the one he dreaded most, the long bathroom conversation.

Three minutes of people talking in one room is boring unless they move, and every move has to be tracked. If you don't reconnect each character with their position for that specific line, the model loses them. Julie had it nailed in the first week, with Henry giving notes on deliveries and blocking, and the shot where the camera goes through the window was her idea. When a generation got the taps wrong, they fixed it with video-to-video passes and a mask.

The scene that really fought back was the escape between the giant creature's legs. It can't be done in one 15-second generation, even multi-shot, so it has to be built from pieces, and every piece has to agree. If Artie is leaning left at the end of one shot, he has to still be leaning left at the start of the next. The legs have to be in the right place, and the yard has to match. He spent a couple of days on 10 or 15 seconds of it.

“I'm fine tuning the prompts manually, but I'm not writing the first draft myself. It's ridiculous. Don't do that.”

Henry Daubrez▶ 37:13

Even with the agent drafting prompts, he calls the process alchemy: you keep throwing things at it until something works. Chris pointed out the irony. An intense action beat can now be one 30-second Seedance 2.5 shot, while two kids talking in a bathroom is the hard part.

Henry tried Seedance 2.5 when it came out, but it changed the polish of the image, closer to the early CGI shows of the 90s, and too different from everything he had made so far. So Chapter 2 stayed on Seedance 2.0, with a move to 2.5 possible next time. And even with long generations on offer, he asks for short ones on purpose.

“If I ask for a delivery of a line in 15 seconds, the delivery is not going to be snappy, because it's going to stretch over 15 seconds.”

Henry Daubrez▶ 40:00

Clean enough to hand over.

Chris asked Henry to share his screen, and he started where every episode starts: the script. Around 15 pages for 15 minutes, which he says was an insane amount of work and wants to trim next time. Some assets already existed from earlier episodes, but without a script he doesn't know what he is making.

“As long as I don't have a script, it's hard to know what I'm going to do.”

Henry Daubrez▶ 41:59

The script keeps moving in production. Lines change when a delivery or the chemistry doesn't work, and the walkie-talkie ending belonged to the next episode until he pulled it forward as a better hook. Each scene is marked with the environment it uses and every line is broken out, because this time someone else had to pick it up. He admits he wouldn't have been that tidy alone, and thinks it is a healthy habit anyway.

Andrew asked whether working this way is easier. Henry's answer was about the word work. Squeezing a whole team into one person with these tools feels like magic to him, but it is still a lot of manual work, and he knows anyone in traditional animation will push back on calling it work at all.

“It's still work, even if I'm having fun. It's still work. It's just not their work.”

Henry Daubrez▶ 44:39

A separate agent for every job.

In the invideo project, each agent has one job, a habit he admits he doesn't keep as strictly as he should. One looks after the vertical fillers he posts between episodes, which need a different style of prompt to feel as dynamic. Chris pointed out that Gossip Goblin had described the same split between his vertical and widescreen work.

The fillers exist because releasing every four to six weeks isn't enough to stay visible: stop posting and you stop existing. But Henry won't make filler for its own sake. The latest had Artie fighting in the yard, only to wake up in class with everyone laughing at him, a hint of what's coming that you can skip without missing the plot.

“Even on the fillers, I'm trying to have a short story which is going to be between 30 and 45 seconds, which is still having a start and an end.”

Henry Daubrez▶ 48:10

Back in the main agent, his method has become strict. He used to just ask it to generate and got a lot of things he didn't want. Now he uploads the bible first, characters, environments and the script, and works shot by shot. If a scene is too ambitious he pastes in only that stretch of the script. And before anything generates, he asks the agent to show him the prompt and the references it will use, with a 15-second cap on each shot.

Moon, no moon.

Then the asset folder. Every character has a master sheet, fixed in Photoshop where needed, and then a sheet for every state they appear in. Artie exists clean, dirty, dirty with headphones, and dirty with and without the sword on his back. Using a separate reference for something like the Walkman lets its size, position and even its design drift, so each state gets its own clean sheet.

Objects get their own sheets too, and environments come in every lighting state the story needs, with the moon and without it. The agent can still grab the wrong sheet, which is exactly why he asks it to list the references first.

The trick Chris singled out came next. As characters move, the model makes its own decisions about where they end up, and feeding it the original references again can reset the angle or the position. So Henry takes a screenshot of the shot he is in and gives it to the model, not to continue the clip, but as context for a new angle, together with his references.

“It's just telling the model, this is where he is now. That's what's behind him. That's what he looks like.”

Henry Daubrez▶ 53:16

To grab those frames faster, he vibe-coded his own extractor in Google AI Studio: give it a video and an interval, and it pulls frames at 1080p to download. Chris linked it to Albert Bozesan, who built his own tool for Magnific in the previous session. Henry's own view on sharing all this was modest.

“I like to think that my value is on what I want to say and how I want to say it.”

Henry Daubrez▶ 54:22

About $1,000 a minute.

Lara asked how long an episode takes. Chapter 2 took around six weeks. Henry works from about 8am to 4pm, stops when his son gets back from school, then picks up again for a few hours in the evening. Julie worked part-time around a very young child and a holiday, so between them it came to seven or eight weeks of work.

“My best guess so far is around $1,000 a minute finished.”

Henry Daubrez▶ 56:44

That isn't nothing when a 15-second generation costs so little, and invideo's team tops up his credits when they run low. But the generations are only part of it. There is still a lot of cutting, pasting and going back and forth. He also admitted to a self-inflicted problem: on earlier shorts, a shot facing the wrong way could simply be flipped and nobody would notice. These characters have details that give a mirrored shot away, so that trick is gone.

Then the Premiere timeline: 15 minutes with overlapping voice tracks, sound effects, ambience, music and Socket's radio crackle, which comes from stock libraries. Henry crossfades between layers and lets the sound of one scene bleed into the next at a lower volume, so places feel connected. In one moment, a line heard from above is borrowed from an earlier clip, duplicated and pushed back in the mix.

For Artie's voice clone, he regenerated a 15-second shot of the character talking in a studio with no other sound, to get the cleanest possible lines to train on. And for quick fixes he often goes back to Google Flow, which he knows well after a year as Google's filmmaking resident, then into Photoshop to clean up assets or build their different states.

Reinforce, then reinforce again.

Flory and Zakariah asked where the characters were made. It starts in a sketchbook: Henry draws every character before the design work begins, even if the final look ends up far from it. Artie's first draft came from Midjourney, was fixed in Nano Banana, then cleaned up in Photoshop.

Then he went backwards. Nano Banana is a great model, he says, but it tends to pull a design towards an average, so he took its version back into Midjourney. He runs the image through /describe to get Midjourney's own reading of it, then combines that prompt with his style references and personal profile.

“Each model reads things differently. That's why I want to do describe, because Midjourney reads things differently than the Banana.”

Henry Daubrez▶ 1:03:07

Chris called it reinforcement at every stage, and the reason this is the best animation the show has seen: skills from before AI and after it, stacked on top of each other. The hosts were blunt about the myth this kills. It goes from AI to Photoshop to Premiere, with hundreds of generations along the way. You don't write a prompt and get a film.

Make something worth watching.

Mary, one of the show's regulars, asked where most AI creators fail when they try to scale a short concept into something longer. Henry sees a new wave of creators who are there purely for the social media game, churning out 15-second shots, and people starting to confuse a following with knowing what you're doing. After five years in the space, back to when you had to put voice-over on everything because nothing could talk, that annoys him.

“What annoys me, what triggers me these days, is there's less effort being put towards storytelling.”

Henry Daubrez▶ 1:07:23

He knows the trade-off. Distribution is still what makes you exist, and he could grow much faster making 20-second sequences. Chris shared an article he had read that morning: young viewers' average watch time on long-form YouTube has grown from about eight minutes four years ago to 20 minutes plus. Andrew thinks the people who say attention spans have dropped are missing one word.

“I think they're missing the word initially.”

Andrew Davis▶ 1:09:45

Then Henry offered some hope to anyone without a big following. Over the last six months to a year, studios and streamers have been reaching out to him, and the people who sign off the money aren't fooled by follower counts any more. The hosts added that views can simply be bought: 500,000 YouTube views for under $1,000.

“If you build something which is worth watching, which shows that you understand how a film is being put together, even a short film, eventually I think there's a lot more chance that people will find you, the ones that matter.”

Henry Daubrez▶ 1:12:01

Danny asked about chain of title, and the human precision generative workflow definitions drafted the week before. Henry is in active conversations with studios and with people on the legal and procurement side. For social media he can work the way he does now. For streaming or cinema, buyers expect a clean chain of title.

“Someone has to be able to prove at the very beginning that you or an individual created the asset you're using.”

Henry Daubrez▶ 1:14:39

He is developing a feature film, and every studio he has talked to wants the same thing: an artist who draws the concept, even if it copies an AI design, and signs a release, so every asset traces back to a person. It is one reason he sees the feature as a hybrid project, with a studio and visual artists involved. He was clear it isn't legal advice.

On the law itself, his reading is that using AI doesn't rule out copyright. The case people cite involved a body of work a machine generated from A to Z with no human input. Generate, then edit, write and direct, and you have a much better case, though it is still decided case by case. Aaron added in the chat that it also depends where in the world you are, and Chris's advice was to keep a log of every project as evidence.

Nine episodes, three seasons.

Henry made the soundtrack in Suno, generating to the length he needed and asking for variations on the same theme with different instruments, so he could place them through the episode and edit how the music builds. Chris passed on a tip he learned from Issa: write the video scene into the Suno prompt, with timestamps, so the music knows when the action is coming.

“Forget it just being a music prompt. Literally tell it the video scene.”

Chris Branch▶ 1:22:59

What about Seedance 2.5 adding music nobody asked for? Chris said he has tried every suggested prompt, and that cinematic scenes always get scored while animation often doesn't. Henry's partial fix is Adobe Podcast. It isn't perfect, since it is really pulling one voice out of a single stream, but it can lift light music out from behind dialogue.

Asked how many tools are in the stack, Henry counted out loud and landed on 14ish. Chapter 2 had wrapped that day. After a few days to breathe, the next episode would take another four to six weeks, and he was already writing it with a shorter runtime in mind, because those 15 minutes in six weeks were a rush.

“I think I would have gone faster if I had done anything realistic. I think animation is always kind of a tricky one.”

Henry Daubrez▶ 1:26:07

The good news: invideo has partnered with him on the whole season, so the bandwidth is there. Nine episodes are planned, counting episode zero, with an outline for three seasons. He asked people to follow him on LinkedIn first, then Instagram and YouTube. And Alex's last-second question about the mannequin behind him gave the show its ending: it wears his wife's hand embroidery, Kitsune fox and all.