ByteDance World Model · Krea 3, Sora, OmniHuman 2.0 and LTX-2.5

Quick Summary

• ByteDance demonstrated an unnamed interactive world model at Douyin's Wonderful Night gala. The public evidence is a stage demonstration, not a released product: there is no official product name, access page, price, benchmark or launch date.
• The most striking demonstration reconstructs new viewpoints around a scene from Stephen Chow's King of Comedy using an authored camera path. It suggests a future in which filmmakers direct a navigable scene instead of accepting one generated camera move, but the footage does not establish production reliability or frame-accurate fidelity.
• OmniHuman 2.0 is available through BytePlus as an API. The official page lists image, audio and optional text as inputs, 30 to 300 seconds of output, camera and body movement, multi-person scenes and a price of $0.07 per second. A five-minute generation therefore costs $21 before retries.
• Krea 3 is officially live as of October 15, 2026. Krea positions it as an image editor first: native 2K output, most edits under eight seconds, up to six reference images, $0.06 per 2K edit and $0.03 per 1K edit.
• Krea's speed, price and comparison figures are first-party measurements. Theoretically Media's tests show useful grain cleanup, localized wardrobe changes and first-frame to last-frame continuity, while complex multi-element cinematic assembly still looks weaker than the strongest general image models.
• The claim that Sora is returning remains a report, not a launch. The Information says a startup founded by Jeffrey Katzenberg, former Sora research lead Bill Peebles and Sujay Jaswa is discussing access to Sora. No agreement, product name, consumer release or delivery date has been announced.
• LTX sponsored the video segment about its post tools. The open documentation confirms Alpha Gen beta mattes, SDR-to-HDR reconstruction, native HDR workflows and tiled 4K or 8K refinement. These are generative workflows, so synthesized detail and reconstructed range must be checked against the source.
• World Gacha is usable now and free. Its official page says the browser tool creates fictional regions locally from random numbers, formulas and simulation, exports PNG, SVG, GLB, FBX, OBJ, heightmaps and JSON recipes, and permits commercial or non-commercial use without credit.
• The practical workflow is selective: use Krea 3 for controlled still edits, OmniHuman for licensed avatar performances, World Gacha for previsualization and LTX for experimental finishing. Treat the ByteDance world model and any Sora revival as watch-list items until access and repeatable tests exist.

Credit, evidence date and the status of every claim

The embedded video is by Theoretically Media and was published on October 15, 2026 at 2:47 p.m. Pacific time. This JQ VISIONS article was published on October 16 and checks public evidence available through that date.

Available now: OmniHuman 2.0 through BytePlus, Krea 3, the LTX-2.5 open workflows and World Gacha. Public but unfinished: Alpha Gen is labeled beta. Demonstrated but unreleased: ByteDance's unnamed world model. Delayed hardware: Pico Space Pro is still officially expected in the fourth quarter of 2026. Reported negotiation: a filmmaker-focused startup may receive access to Sora.

That status map matters because a compelling stage clip, an API, an open checkpoint and a press report carry different kinds of evidence. The article uses official documentation for current specifications, attributes creator observations to the video and keeps forecasts clearly labeled as analysis.

OmniHuman 2.0 · Five-minute avatar performances through an API

OmniHuman 2.0 takes one source image, an audio track and an optional text prompt. BytePlus says the model coordinates lip movement, emotion, gesture and camera behavior with the rhythm, prosody and meaning of the audio. The official product page also shows non-human subjects, stylized characters, multi-person scenes and prompt-directed object or camera changes.

The commercial change is easy to quantify. BytePlus lists $0.07 per generated second and advertises 30 to 300 seconds of output. Sixty seconds costs $4.20, three minutes costs $12.60 and five minutes costs $21 for one attempt. The old $0.12 figure remains on the page for the earlier tier, so the new rate is about 42 percent lower.

Theoretically Media has not independently tested 2.0 because access was region restricted in the creator's location. The examples in the video are ByteDance or BytePlus demonstrations. Tim notices improved camera movement and performance compared with older talking avatars, but also points to teeth, exaggerated acting and imperfect articulation as visible weaknesses.

The best near-term use is patchwork: a licensed character delivering a specific line when a general video model keeps changing the dialogue. Do not read API availability as permission to clone a person's face or voice. Record consent, permitted territories, duration of use and deletion requirements before uploading identity assets.

ByteDance's unnamed world model · What the gala actually showed

At the October 11 Douyin and Beijing TV Wonderful Night gala, filmmaker Pan Tianhong demonstrated a ByteDance world model that generated explorable scenes, accepted mouse and keyboard navigation and responded to new instructions inside the environment. Reports from the event describe prompted objects appearing and characters interacting with them.

The cinema demonstration is the important part. A scene from Stephen Chow's King of Comedy appears to be reconstructed from new camera positions along one drawn path. Theoretically Media interprets visible interface elements as environment planes, a camera trajectory and possibly a Gaussian-splat-like representation. Those interface readings are informed observations, not confirmed architecture.

Bloomberg separately reported in September that ByteDance was preparing a real-time spatial video model. It is reasonable to connect that report with the gala, but ByteDance has not publicly confirmed that the demonstrated system is the same project. It has also not published the model name, access terms, technical report, benchmarks or release schedule.

Why re-shooting a scene matters more than another text-to-video demo

Most current AI video workflows generate a flat clip. If the camera is wrong, the creator usually regenerates the shot and risks changing the actor, blocking, light or environment. A spatial representation could instead let a director preserve the dramatic moment and author a new camera path around it.

The promise is closer to virtual production: establish a world, block the action, move a camera, change a prop and render coverage. That would connect video generation, 3D tools and interactive world models. It could also make continuity and editorial coverage more controllable than repeatedly prompting independent clips.

The unresolved tests are demanding. A production model must preserve identity, occlusion, reflections, background geometry, lip sync and motion across every new view. It must also expose licensing and provenance for source footage. One authorized Stephen Chow demonstration cannot prove that the system works on arbitrary films or that users may re-stage copyrighted scenes.

Pico Space Pro · A possible home for spatial generation, not a confirmed bundle

Pico, ByteDance's XR subsidiary, postponed the Space Pro launch from September to the fourth quarter of 2026. Road to VR reports that the device cleared the FCC and that Pico expanded a closed early-access program, but Pico has not announced a final public date or official price.

The video suggests the world model may eventually connect with Pico. Bloomberg also described possible applications across Pico and Douyin. That connection is plausible, but there is no public confirmation that Space Pro will ship with the gala model or that creators will receive the same camera-path tools.

Treat all price estimates as unconfirmed. Businesses evaluating the headset should wait for Pico's own SKU, region, warranty, storage, compute and software terms rather than budgeting from leaks or reseller placeholders.

Krea 3 · The official release is an editor first

The video was recorded around early access, but Krea's official October 15 release changes the status. Krea 3 is live on Krea and through several API partners. The company lists native 2K editing, up to six references, most edits under eight seconds, $0.06 per 2K image and $0.03 per 1K image.

Krea says the model is trained to change the requested region while preserving products, people, labels and composition elsewhere. That is more valuable for production than a beautiful one-off generation. A campaign team needs the bottle, logo, face and layout to survive a material swap or new background.

Krea's comparison chart reports 33 seconds for Nano Banana Pro and 90 seconds for GPT Image 2 against under eight seconds for Krea 3. Those measurements come from Krea's own jobs and pricing assumptions. They are useful first-party data, not an independent benchmark across every provider, queue and prompt.

Krea 3 tests · Grain cleanup, local edits and continuity frames

Theoretically Media first removes the familiar grain pattern from a GPT-generated portrait and improves skin texture. The result is visibly cleaner. A second test changes a recurring character's clothing, and precise edit mode lets the creator mark only the region that should become business attire.

The strongest workflow example builds a new last frame for a diner scene: the opening image shows an agent with coffee, while the edited frame adds a waitress and pie without wildly changing the table. Those two anchors are then used for first-frame to last-frame video generation. This is a practical way to direct the destination of a shot instead of describing the entire motion in text.

The weakness appears when several references must be combined into a dynamic cinematic composition. Krea 3 follows the idea, but the creator prefers GPT Image 2's visual drama despite its grain. The conclusion is specific: Krea 3 looks like a strong controlled editor, not automatically the best generator or compositing model for every shot.

LTX-2.5 post tools · Useful, open and explicitly sponsored

LTX sponsored the section from 10:11 to 12:45. The sponsorship does not make the tools irrelevant, but it changes the evidence role. Feature availability can be checked in the public repositories and documentation; quality claims still need independent footage and measurement.

Alpha Gen beta creates a frame-aligned grayscale matte from a source video. LTX documentation warns that the model selects what it interprets as the main subject, does not accept a prompt or mask for subject selection, and is most predictable at 121 frames or fewer. Treat the output as a starting matte, not a promise of production-ready edges.

The SDR-to-HDR workflow reconstructs an HDR result and can output scene-linear ACEScg EXR plus an HLG master. LTX explicitly distinguishes this from preserving an existing HDR source. Native HDR workflows keep an EXR-based path for image-to-video and inpainting. That distinction should survive marketing shorthand.

Tiled Fusion workflows offer Full HD, 4K and 8K paths with a detail refiner. The documentation calls the process generative and says it synthesizes detail rather than performing pixel-accurate restoration. For faces, products, typography or evidence footage, compare every important region with the source before delivery.

Sora returns? · What The Information reported and what it did not

OpenAI's own Sora page says the consumer product became unavailable on April 26, 2026. On October 12, The Information reported that a stealth startup focused on filmmakers was in talks to access Sora, modify it and sell software or services built around it.

The reported founders are Jeffrey Katzenberg, former Sora research lead Bill Peebles and former Dropbox CFO Sujay Jaswa. The report says OpenAI would receive equity if a deal were completed. It does not establish that a deal has closed, that model weights will transfer, that the product will be called Sora or that consumer access will return.

The accurate headline is therefore possible professional reuse of Sora technology. A studio should not put it on a roadmap, promise it to a client or compare its price until the parties announce a completed agreement and actual access terms.

The OpenAI recorder case · Keep the allegation precise

The U.S. Department of Justice says an OpenAI engineer was arrested on October 9 and charged after a concealed audio recorder was discovered in an executive conference room. The complaint describes an account associated with the defendant's spouse that submitted prompts about building, installing, hiding and remotely accessing a recorder.

The video says the how-to prompts came from ChatGPT. The public Justice Department release says only a chatbot account. Because the specific service is not established in that document, this article does not repeat the stronger attribution. The charges are allegations, and guilt must be proven in court.

This story is not evidence about the quality of any creative model. Its relevant lesson is operational: do not put credentials, private schedules, confidential meetings or unauthorized recordings into an AI workflow, and do not treat an assistant as permission to perform a prohibited act.

World Gacha · A free local city generator with Blender exports

World Gacha is the most immediately accessible tool in the episode. The official page credits MIYADAIKU as the maker. The video says Theoretically Media discovered it through Sharp, but the project credit belongs to the creator named on the tool page.

The browser tool generates nine kinds of fictional regions using random seeds, formulas and erosion simulation rather than external map data. Users can change terrain, water, land use, roads and camera position, then export PNG, SVG, GLB, FBX, OBJ, a 16-bit heightmap, a map image, a share link or a JSON recipe.

The creator says generation runs inside the browser and data is not sent elsewhere. Exports may be used commercially or non-commercially without attribution. For a film workflow, save the seed and recipe, pick camera positions, export a lightweight GLB or FBX, validate scale and normals in Blender, then use the scene for storyboards or layout rather than assuming it is a finished production environment.

A practical stack you can test this week

Start with a licensed still, a short voice recording and one synthetic environment. Use Krea 3 to create a controlled wardrobe or set change while preserving the face and product. Save the original and edited anchors. If the character must speak, test a 30-second OmniHuman performance and review articulation, teeth, hands, gesture intensity and camera behavior.

For spatial planning, generate a World Gacha location, save its recipe and export the scene to Blender. Use it to choose focal length, blocking and camera paths. Do not claim that this reproduces ByteDance's unreleased world-model workflow; it is a current, controllable substitute for layout and previsualization.

Finish a short clip with LTX only if the post operation solves a measured problem. Inspect Alpha Gen edges over a contrasting background, compare SDR-to-HDR highlights against scopes and source intent, and review refined frames at 100 percent for invented texture. Keep a conventional matte, grade and upscale fallback.

The acceptance gate is simple: the edit must preserve protected regions, the performance must use authorized identity assets, the 3D export must reopen correctly, and the post result must survive a blind comparison against the original workflow. If it fails any gate, the experiment remains a prototype.

What to watch next · Evidence that would change the verdict

For ByteDance's world model, look for an official product page, access program, technical report, maximum world duration, camera controls, export formats and tests on material beyond the gala. For Pico, wait for official price, region and shipping details.

For OmniHuman 2.0, the next useful evidence is independent testing across accents, singing, profile views, multiple people and long clips. For Krea 3, look for repeated edits on typography, packaging, architecture and faces with untouched-region difference tests.

For Sora, only a named product or completed partnership changes the status. For LTX, look for independent comparisons of matte quality, HDR reconstruction and identity preservation. Fast-moving tools deserve short evidence windows and dated language, not permanent superlatives.