Google's new Genie 4 can take a single real photo and spin it into a world you can walk through for two minutes. In this episode of Denoised, we tested it on Hearst Castle, a Culver City street and a Western saloon with a whale in it. We also listened to Block's Dessa voice model, argued over IATSE's new labor and technology intelligence officer, and broke down the rule change Epic Games made to its MegaJam after AI backlash.
Jump to:
00:31 Genie 4's two-minute interactive worlds
08:01 Turning a Hearst Castle photo into a 3D space
14:12 Dessa, Block's voice model
22:38 IATSE's new AI officer
31:43 Epic's MegaJam AI backlash
Quick Take
Every tool in this episode got closer to real. Genie 4 builds believable rooms from one photograph and Dessa breathes and laughs like a person, yet Genie 4 forgot whole rooms the moment we turned around. Meanwhile, IATSE wants members to have a say in workflow, and Epic rewrote its game jam rules after AI backlash.
What We Tested: Genie 4 Builds Two-Minute Worlds From Prompts and Real Photos
We started with silly prompts and finished with real locations. Genie 4 is still an early beta in Google Labs, and Google DeepMind built it with live text steering and native audio; music played softly through our sessions.
What changed from Genie 3:
- Two-minute sessions. Genie 3 gave about 60 seconds of mostly persistent world; Genie 4 doubles that.
- Three ways in. An environment prompt, an optional character prompt, or a starting image.
- 720p output. Detail is much better, down to the chairs.
- Live events. Type an event mid-session, pick a suggested action, or hit "I'm Feeling Lucky" to drop something random into the scene.
- First and third person. Explore as the camera or generate a character to walk through the world.
- Recording and API. Sessions record to a downloadable video, and Google is rolling out an API for the first time.
Prompt tests got weird fast. We asked for an 1870s saloon in Yuma, Arizona, with a whale as the cowboy, and got a humpback whale in the bar being offered drinks. When we turned around, the whale and the entire bar were gone.
The event controls worked, loosely. A "saloon showdown" suggestion opened the doors without a showdown; a second auto-generated prompt produced a guy with his gun drawn, and typing "someone falls through the ceiling" did exactly that.
A suggested boat-at-a-pier prompt produced what Addy read as '90s Hong Kong. When we turned back, the two boats and the person we had walked toward were gone.
Real photos held up, for a while. Fed a photo from Hearst Castle, Genie 4 generated an entryway that was never in frame and still felt right for the building. The table stayed consistent when we came back, a suggested prompt made books float off the shelves, and we could grab a ladder and try to push it.
Addy's reaction: "It's cool that it's from a real photo. It's extrapolating all that from a single photo."
Then it broke. The room we tried to loop through appeared to have two doors; inside, the second door did not exist, and stepping back out landed us in a completely different space.
A Culver Theater photo held together at first, but moving down the street turned it into a random block, and when we went back the theater was gone.
Addy's take on the soft detail: as time increments, the model has to track every element and its position, so memory demand climbs fast. Shortcuts to keep memory low would explain the low texture detail.
Addy's counter to the skeptics: try building the Hearst Castle room in traditional CG. You would model every bookshelf, place every book and match the floor texture; the result would be better, but he estimates a month of work for a 5 to 10 second scene.
Where we can see it fitting production:
- Quick previs iteration on locations you already have photos of.
- Two-minute location scouts where you explore a picture with no real-world consequence.
- A starting point for splats. Because every session is recorded, a world you like could be turned into a splat.
Joey's argument on why Google is building it: he thinks it ties into world understanding that could ultimately feed robotics and self-driving. Addy pictured it tied to Google Maps and the imagery Google already holds, and said the live prompt bar is begging for an agent that steers the experience and acts as the world's logic.
What We Explored: Dessa Sounds Human Because Block Trained It on Audio
We listened to sample conversations from Block's Dessa voice model, and the breaths, laughs and pauses are what stand out. Block, the Jack Dorsey company that runs Square, says it trained Dessa only on speech; as Joey pointed out, many other voice models take a multimodal approach instead. Dessa is a preview and not open yet.
Addy's explainer on full duplex: true two-way conversation, where both people can talk over each other, is what an analog phone call in the '90s felt like. Zoom and ChatGPT voice are not there; the assistant takes a second or two to process, and people on calls stomp over each other and then defer.
The disclosure problem: if an AI agent gets you faster service from a business, do you want it to sound fully human, or do you want to know you are talking to a machine? Addy predicts voices will get very natural, people will find it unsettling, and companies will swing back to a clearly synthetic sound.
Joey's argument: fictional computers like HAL and Jarvis have distinct computer voices and never try to imitate a person.
The risks are concrete. Addy got a callback from his internet provider and could not tell whether it was AI; it sounded natural but kept talking without the cues a person gives. Better voices also mean better telemarketing, impersonation and scams aimed at elderly people with limited hearing.
Production uses we see:
- Dubbing AI characters. Addy's idea: generate a performance from a transcript, then re-voice it with a model like Dessa so synthetic characters sound more natural.
- Lip remapping. If Dessa handles multiple languages, pairing it with lip-sync tools from other companies is the next step.
- Retail phone agents. Joey's guess at Block's motive is offering Square merchants a natural-sounding agent that answers every call.
What We Debated: IATSE's New AI Officer and the Case for Analog Workflows
IATSE named Jillian Arnold its first labor and technology intelligence officer, tasked with working out what AI means for reskilling and educating the trades the union represents. Arnold comes from cinematography and has a long history with the union.
In an interview with The Hollywood Reporter, Arnold said members can talk with an employer about AI before employment begins, and she wants them to be able to communicate if they feel the right approach is "what we could call an analog workflow." She added that some of those analog workflows are still considered more economical and produce a unique, copyrightable result.
Joey's argument: the job is translating hard-earned taste into generative tools. A cinematographer's lighting sense can direct image models like Nano Banana Pro, and a production designer's eye carries into working with reference images and video models.
Addy's counter: reskilling only works with people who want it. "For the person to be reskilled and be adoptable in this new era, they have to be willing participants," he said. He sees a culture shift inside IATSE as the precursor to any upskilling, and thinks Arnold could help drive it while many members feel threatened and the industry is already hurting.
Joey's read on the analog line: some tasks are faster when someone who knows Blender or Unreal Engine does the work directly. VFX often stays hybrid, with generated assets next to traditional compositing and green or blue screen where pipelines are established.
Addy's concern: keeping every member employed and keeping the industry profitable are different goals. If the union puts up blockers, he worries productions will hire non-union, and if AI cuts a seven-day shoot to three, the set builds, grips, lighting crews and drivers for those days go away.
Where we landed: production has been leaving LA for economic reasons unrelated to AI. The optimistic case is lower budgets producing more and riskier projects, and smaller G&E teams opening the door for a best boy to gaff other shows.
What We Questioned: Epic Changes Its MegaJam Rules After AI Backlash
Epic Games announced the AI MegaJam, drew criticism, and changed the rules. The jam offers $50,000 in cash prizes and gives teams of up to five humans, plus as many AI agents as they want, four days to build a playable game in Unreal Engine or Unreal Editor for Fortnite.
As we read the change, AI appears to now be optional rather than required.
Addy's take: it should always have been optional, and he expects teams that use AI to make the better games. He connects the jam to a hard pivot at Epic that followed what he described as a 1,000-person layoff.
The Roblox factor: Addy sees the jam as part of Epic's push into user-generated Fortnite content, similar to Roblox, though he doubts Epic can match Roblox's volume of games and expects Roblox to add vibe-coding tools to Roblox Studio too.
Joey's argument: optional AI is the better outcome, because now we get to see what the winning team used. The rules bar any work before November 12, which addresses the worry that teams could show up with a game already built.
Why game developers push back harder: Addy thinks the game industry has been in decline longer than film, pointing to layoffs at EA and Xbox and Netflix shutting down gaming divisions. After years of that, being told AI will do your job lands badly.
Bottom Line: The Tools Are Getting Real Faster Than the Rules Around Them
Each story this episode put more realism in front of us and raised the question of who decides how it gets used.
- Genie 4. A single photo now becomes a walkable two-minute world we can see being used for quick previs and scouting, until it forgets where you have been.
- Dessa. Block's speech-trained voice breathes and pauses like a person, which could help dubbing and raises real disclosure and scam questions.
- IATSE. Arnold's appointment puts members' input on workflow, including analog, on the table, and Addy argues willingness has to come before reskilling.
- Epic's MegaJam. With AI apparently optional, the winning entry will show which tools teams choose on their own.
Each tool arrived with more realism than control, and the open questions are about who decides when it gets used.
Links from This Episode
Tools & Platforms
News & Analysis
WATCH THE FULL EPISODE ON YOUTUBE →
