ByteDance's World Model Re-Shoots a Real Movie, Krea 3 Is an Editor First, and Sora Might Come Back
ByteDance showed an unnamed world model that re-shot a Stephen Chow scene from one drawn camera path, on phone footage from a Douyin gala. Also this week: OmniHuman 2.0 (API only, not in the US), Krea 3's beta as a fast image editor that will actually edit Flamethrower Girl, The Information's report that Sora may return through a Katzenberg and Peebles startup, Pico Space Pro, and World Gacha, a free city generator that exports to Blender.
ByteDance put a world model on real movie footage, Krea shipped a 3 that’s an editor more than a generator, and Sora, of all things, might be coming back. I never thought I’d say that. The full rundown’s in the video; here’s the week in order.
Three hedges before we start. The world model is unnamed and unreleased, and everything I’ve seen of it is a phone pointed at an LED wall. Sora’s return is a report of talks, not a deal. And OmniHuman 2.0 is region-locked for me, so that section is ByteDance’s examples, not my tests.
OmniHuman 2.0: ByteDance’s lip-sync avatar is back
OmniHuman is ByteDance’s avatar lip-sync model, and 2.0 went live October 15 through BytePlus: more camera movement, more emotion, better lip sync than your standard avatar. BytePlus’s own list: lip sync that follows tone and emotion, clips up to 5 minutes (from 60 seconds), about 40% cheaper, native 720p, and camera, gesture and multi-person control by text prompt. It’s image plus audio in. You can’t drive it with video.
I don’t think it looks bad. We’ve come a long way from the bobblehead hand-motion days of avatar tech. The teeth are still a problem, and while I don’t read lips, it doesn’t feel like it’s hitting all the diphthongs. (Diphthongs: an incredible album name for my fake punk band, the Gaussian Splats.)
One thing I’ve noticed about ByteDance’s examples is that they never look as good as what we get once we have the model ourselves, and this batch is a bit over the top. Fake Karl Urban really needs to chill out on the coffee. That said, for what is essentially an avatar, it clearly does cinematic camera. It feels a little south of Seedance 2 and 2.5, and you might not be wowed, but tools like this come in very handy for patchwork in an AI film, when you need one specific line of dialogue to stay the line you wrote. Seedance will often change the line on you. OmniHuman won’t.
It’s API only for now (“contact sales”), and BytePlus isn’t available in the US, which is the region lock. I checked with the Dreamina team: my read is that it eventually lands on that platform, but nothing is announced, and Dreamina’s avatar tool still runs 1.5.
First look: ByteDance’s unnamed world model
Now the more exciting ByteDance news. At the Douyin Wonderful Night gala in Beijing on Sunday, October 11, ByteDance showed its world model for the first time. It has no name and no release date, the presenter was a creator (Tim Pan of MediaStorm, a different Tim), and what we have is bootleg footage of an LED wall. With all of that said, it looks really good.
Initially it’s the standard world-model thing, with one difference I was really happy to see: mouse control. Every other world model is WASD and arrow keys. You can also prompt objects into the world and have the characters react to them; a moon pops into existence and the character looks up at it.
The part that had me doing the Leo-pointing-at-the-screen meme was camera re-tracking inside a scene from Stephen Chow’s King of Comedy. That shouldn’t be a surprise: Chow’s company signed a licensing deal with ByteDance’s Volcano Engine back in June, covering King of Comedy, God of Cookery and CJ7 for AI video (I covered it at the time). The demo takes the environmental details and the performances of the original shot and gives you new coverage of it: a dozen camera positions, previewed by drawing a single camera path, with Chow’s permission obtained in advance.
It’s hard to tell from phone footage, but the UI in the corner looks like it has 3D controls, and the second shot might be a Gaussian splat. A portal-effect demo gives another peek at the interface when it cuts wide: three environment planes, and what I presume is the camera track drawn through them. I doubt that’s the final UI, but it’s an interesting way of looking at it.
All of this opens the door to use cases beyond the current world-model template of walk around an environment for 60 seconds while things decohere. More than that, the modalities are coming together: video generation in Seedance, 3D asset generation, world models. I expect 2027 to be the year that brings it all together. (World models had a loaded week in September too.) Bloomberg reported in September that ByteDance is building a Seedance-based real-time world model for Pico and Douyin; ByteDance hasn’t confirmed that’s what we saw.
Pico Space Pro: ByteDance’s Vision Pro?
Which may have something to do with ByteDance’s upcoming Pico Space Pro headset, its version of the Apple Vision Pro. It isn’t out yet: the September launch was cancelled and it’s now slated for Q4 2026. There’s no official price. The leaks point to roughly $2,400 to $2,600 against the Vision Pro’s $3,699, so cheaper by about a third, not half. I’ll keep an eye on it.
Krea 3: an editor more than a generator
Back on home turf, Krea has clearly been drinking fake Karl Urban’s coffee: Krea 3 is here, five months after Krea 2. It’s a beta as of October 15, with the final release due the following week.
The important part is that Krea is positioning 3 as an image editor more than an image generator. Generation is coming (more below), but the pitch is speed on edits: Krea says most edits come back in under 8 seconds at native 2K. It’s actually pretty good.
As a straight generator, it’s whatever gets the job done. Four Man in the Blue Business Suit frames walking down a busy city street are fine. Krea seems to know this: try to generate an image in Krea 3 and it asks whether you’re sure. I think that’s because a lot of Krea 2’s bells and whistles aren’t here yet.
| In Krea 3 today | Coming |
|---|---|
| Precise mode: paint boxes on what changes, the rest stays put | LoRA generation (waitlist) |
| 2K edits, most under 8 seconds per Krea | Style transfer, which was a very strong part of Krea 2 |
| $0.03 per 1K image, $0.06 per 2K | The final release, the week after the beta |
As an editor, it does very nice stuff. We all know the grain pattern in GPT Image. Hand that image to Krea 3 with “remove the grain and improve the skin texture” and the result is significantly cleaner.
Something more me-specific, or specific to fans of Flamethrower Girl: any time I try to edit her original Mad Max outfit in Nano Banana or GPT, I get pushed back for inappropriate content. Quick side rant: I get it, she’s fairly cheesecake in that one, but I’ve always considered her a superhero, and there are iterations of Wonder Woman wearing less. Krea 3’s response to an outfit change is: sure. (I know Claude doesn’t do images, but if I tried this in Claude I’m 90% sure I’d get a lecture.) Precise mode, paint a box, “business attire,” and there you go: Business Shark Flamethrower Girl. Ask her for those reports and she’s going to burn down the office.
Where this really comes in handy is first-frame / last-frame work. Take our FBI agent drinking coffee in a Pacific Northwest diner, generate an edit of the waitress delivering the pie, use that as the last frame, and the result is pretty solid. It fixes problems we’ve hit over and over: the water cup that vanishes mid-shot, the insanely sized slice of pie. A lot more control.
Where it falls short: combining multiple elements into one frame for cinematic work. Krea 3 gets the idea and gets the job done, but GPT Image’s version of the same composite looks a lot more dynamic, grain and all. And when you take all those elements into an omni model, it soups them together anyway (there’s a Planet Hell work-in-progress frame in the video to show it).
Not going to be my primary image generator. But it’s very fast, and I think it finds its way into my pocket as my primary image editor.
Sora is coming back?
Sora doesn’t need an introduction; I wrote the obituary. The plug got pulled earlier this year: the shutdown was announced in March, the app and site went dark in April, and the API ended in September, when OpenAI decided it was done with Sora and moving on to much more serious projects, like Dots. I’m kidding. I love my Dot. When I recorded, sora.com was still up and you could still download your assets, though none of the icons do anything anymore. Pretty sad.
In the wake of all that, Bill Peebles, Sora’s research lead, left OpenAI in April. Now, per a report from The Information (October 12), OpenAI is in talks to give an unnamed stealth startup deep access to the Sora technology in exchange for an equity stake. The startup isn’t all that stealthy: Peebles is a co-founder, as is Jeffrey Katzenberg, the former head of DreamWorks Animation, along with Sujay Jaswa, the former Dropbox CFO who co-founded WndrCo with Katzenberg. (On camera I said “the weights and the model.” The Information’s wording is “deep access,” and the weights are unconfirmed.)
| Reported | Not reported |
|---|---|
| Talks to give the startup deep access to Sora, enough to modify it and sell software built on it | A signed deal |
| OpenAI takes an undisclosed equity stake | A timeline |
| The customers are film studios, not consumers | The weights changing hands, or a product name |
So, about as stealthy as the guy who got caught putting microphones in the OpenAI conference room, which is to say, not very. (An OpenAI engineer was arrested October 9 for allegedly hiding a recording device in an executive conference room; per the complaint, the how-to prompts for building it came from ChatGPT. Allegedly. I am not an international spy, but apparently neither is this guy.)
Back to Sora: nothing suggests an imminent release, and it probably isn’t consumer-facing this time. The Information says the audience is a narrow, professional film-industry one rather than a wide release. We’ll see how it plays out. I’d love to see Sora back in action.
World Gacha: free cities you can wander, and export to Blender
Rounding out with a neat little thing from Sharp (@sharp_maker): World Gacha generates fictional regions and towns, builds them out, and lets you wander around in them, down to street level. It’s all procedural, from random numbers, formulas and an erosion sim: no map data, no image AI. Built with Claude Code, per the maker.
On camera I called it open source, and it isn’t. It’s a free browser tool, and everything you export is free for commercial use with no credit required, but there’s no code release or license. It was all in Japanese when I recorded (I’m terrible at reading kanji, hiragana or katakana); an English toggle landed afterward, at ?lang=en. Zoom with the mouse, pinch on mobile, pick from a few render styles, roll another city.
The interesting part for us: you can export the geometry (OBJ, GLB, FBX, plus PNG, SVG and a heightmap). Generate a section, import it into Blender, and do all your Claude or GPT Blender MCP stuff on top of it, or use it to populate a city for references. I haven’t tried the export yet. Worth a bookmark, and if nothing else, worth five minutes of zooming in and out of a city, which is what I’ve been doing.