ByteDance's world model re-shoots a movie scene, and Krea 3 lands

Theoretically Media's latest episode covers more ground than any single news cycle usually supports, and most of it is checkable. So this is the workshop version: every segment in the episode, what the host says about it, and what the primary source says when you open it. Tim Simmons publishes his own notes and chapter list with each episode, and this breakdown is built from those plus the links he provides; we could not retrieve the episode's captions, so where his narration and the documents disagree, we say so and go with the documents.

The episode's five headline items are a ByteDance world model nobody has named, Krea 3, the report that Sora returns through a filmmaker startup, Pico's delayed Vision Pro rival, and a ByteDance avatar model. It also carries a sponsored segment we will identify as one.

OmniHuman 2.0 is back, API only, and not for US buyers

The first three minutes go to ByteDance's image-plus-audio avatar model, out through BytePlus. The host's own note is precise: 720p, API only, and not available in the United States, which is why his own demo is region-locked.

His verdict is the useful part: it sits "a little south of Seedance 2.5", which makes it a tool for patchwork lines rather than hero shots. We covered the launch and the price ourselves in OmniHuman 2.0 generates 5-minute avatar video at $0.07 a second, so this breakdown keeps to what the episode adds.

The world model ByteDance will not name

This is the segment worth your time, and it comes with the most honest caveat in the episode: no name, no date, no press release, and phone footage of an LED wall. What the host says he is watching is a world model shown on Sunday at Douyin's Wonderful Night gala, controlled with a mouse and keyboard, where you prompt an object into the scene and the character picks it up and uses it.

Then the demo turns to cinema. From the film King of Comedy, the model re-shoots a scene from roughly a dozen camera positions generated from a single drawn path, and the host notes that ByteDance had Stephen Chow's permission to do it. His reading of the demo's trajectory is the argument to take away: video generation, 3D and world models are converging, and he dates the convergence to 2027.

The reporting behind the demo is Bloomberg's from September: ByteDance is preparing a real-time spatial video model for its own products. Independent writeups of the gala demo describe a scene generated from a single sentence that you can then move through, and a camera path you can author rather than accept. Treat the demo as a demonstration of direction, not a product, because that is all it is.

Pico Space Pro: the delay, the specs, and the price nobody has published

ByteDance's headset subsidiary has an Apple Vision Pro competitor, and the host gives it about twenty seconds: delayed to Q4 2026, no official price, and leaks of roughly $2,400 to $2,600 against the Vision Pro's $3,699.

The documents are more interesting. Road to VR reports that the headset, previously code-named Project Swan, has now cleared the FCC, which is one of the last gates before US sale. The filings call it a mixed reality computer and describe microOLED displays at 4,000 pixels per inch with an average of about 40 pixels per degree, plus a dual-chip design: custom XR silicon for perception and imaging with a claimed 12-millisecond perception latency, and a separate flagship chipset the company says is more than twice the CPU and GPU performance of the XR2 Gen 2. Pico cancelled the Beijing reveal that was booked for September 2, blamed "a major upgrade to the software experience", and is still saying Q4 2026.

On price, the honest state is this: Pico has published nothing. The host cites leaks around $2,400 to $2,600. The only dollar figure attached to the device in public is a US reseller listing at $2,899, dated February 2027 on an enterprise page, and a reseller is not the maker. The headset is launching into a crowded field, too: Vision Pro, Samsung's Galaxy XR, and Valve's Steam Frame.

Krea 3: an editor first, and the number is eight seconds

The episode calls Krea 3 an editor more than a generator, and the beta-versus-release distinction matters. The host says he is filming against a beta released that day, with the final release the following week. Krea's own launch post says the final model is released and live for its 30 million users, available through fal, Hedra, Together AI, Nous Research, OpenRouter and Replicate. So one of the two statements is out of date by the time you read this, and the model's own page wins.

What the page establishes: most edits come back in under eight seconds end to end at native 2K, at $0.06 an image, up to six reference images per edit, and the model keeps what you did not ask it to change, so a label stays sharp when the scene around it moves. Krea's comparison, measured on its own service between September 24 and October 1, puts Nano Banana Pro at 33 seconds and GPT Image 2 at 90 seconds, at $0.135 and $0.22 an image.

The host's tests are the part you cannot get from a launch post, and they are the reason this segment is in the episode at all: grain removal on a GPT Image shot, an outfit change on Flamethrower Girl, a diner scene built from a first and last frame, and a section of the episode given to where it falls short, with a Planet Hell work in progress as the stress case.

Sora comes back, if it comes back

The episode returns to the report that OpenAI is in talks to let a filmmaker startup use the shelved Sora model, and the host's framing is careful. He raises it twice and does not pretend to know how it ends. We covered that report in Report: OpenAI in talks to license Sora to a filmmaker startup, so this breakdown keeps to what the episode adds.

The arrest, and one detail the episode gets wrong

The episode raises the arrest of an OpenAI engineer on October 9 over a recording device hidden in an executive conference room. The host's own note credits the complaint: per the complaint, he writes, the how-to prompts came from ChatGPT, and he marks it as alleged.

Fortune's account of the complaint describes prompts sent between September 24 and September 30 from a ChatGPT account linked to the wife of Abdulrahman Alfozan, a 35-year-old OpenAI engineer and Saudi national who has been a US legal resident since 2020. The prompts read like a build order: a continuous audio recorder, live listening, download and delete controls, an encrypted SD card, a power-only USB output, a second unit with speech-triggered recording, a third unit, and finally silencing the boot output. Prosecutors say the device was hidden behind a television in a conference room on or about September 28. The complaint does not charge his wife, and it does not list which chatbot was used; OpenAI confirmed to Fortune that the chatbot was the company's own ChatGPT.

Two further details from the documents are worth carrying: about six weeks earlier, the affidavit says, he used an internal OpenAI productivity agent to find rooms seating six or more people for high-powered meetings, and the room he chose was one he had no access to, reserved for executives and their assistants. He faces one count under the federal Wiretap Act, which carries up to five years, and whether the device captured anything is not established.

World Gacha: free cities in a browser tab

The last segment is the cheapest and, for anyone who builds sets, the most immediately useful. World Gacha is a free browser tool that rolls an entire fictional region from a seed. One button produces a town in one of nine region types, a castle town, a hillside town, a skyscraper district, a water city above a waterfall; you can draw on the map to correct it, place a camera anywhere, and export the result as PNG, SVG line art, GLB, FBX or OBJ, along with a 16-bit heightmap. It renders through three.js, runs entirely in the browser with nothing sent to a server, and has an English toggle.

One attribution note, because the two names belong to the same maker. The episode credits the tool to Sharp (@sharp_maker); the tool's own page credits MIYADAIKU of Oyasumi Game Dev. That site's operator handle is sharp_maker, the name on its X account, and MIYADAIKU is the pen name its tools are credited to.

The sponsored segment, labeled

At 9:57 the episode stops for its sponsor, LTX, and the segment is a genuine release rather than an advertisement for nothing: LTX's post-production notes describe six new and upgraded capabilities built for finishing work. Alpha Gen hands you an alpha matte for a shot that keeps soft edges and fine detail, so the hair, the smoke and the partial transparency survive compositing. SDR to HDR turns 8-bit footage into 16-bit EXR, using the original SDR clip as the control signal and writing an HLG master alongside it, which is published as an open-weights LoRA. HDR Native keeps 16-bit footage intact through the edit, and Native Resolution with Refine renders at 4K or 8K and rebuilds detail rather than upscaling it. The weights are open, the workflows run in ComfyUI, and there is a desktop build for people who do not want to assemble a graph. We say it is sponsored because it is, and because you should weigh a host's enthusiasm accordingly.

What to do with the episode

• Watch it if you make anything with generative video, and subscribe to Theoretically Media while you are there. He tests in public, states his limits, and says when something is a demo rather than a product, which is rarer than it should be.
• If you only have five minutes, take the world model segment and the Krea 3 caveats. The first tells you where the next eighteen months are going, the second tells you what a fast editor is worth when you are making ten variants instead of one.
• If you build sets, open World Gacha. A free browser tool that exports GLB and a heightmap is a week of work you do not have to do.

The episode is "ByteDance's World Model Re-Shoots Movies, Sora Returns? & Krea 3 Is Here!" on Theoretically Media. We are viewers of the channel, not colleagues: the work in it is his, and the checking here is ours.