Xmax X1 — Real-Time Interactive Video Model (Product Announcement)

TL;DR

Xmax AI (a Tsinghua-affiliated startup) announces X1, a closed-source real-time interactive video generation model pitched as the “first virtual–real-fusion real-time interactive video model.” X1 takes a phone camera feed plus touch-gesture inputs and re-renders a fused virtual + physical scene at claimed millisecond latency on consumer-grade hardware. The technical description (only available through company blog posts / Chinese tech press, not a paper or report) names the architecture as a frame-level autoregressive DiT with end-to-end streaming re-rendering, plus a unified interaction model that fuses 3D spatial relations with 2D screen gestures. There is no public arxiv paper, no model weights, no benchmark numbers — only marketing material, a demo video, and a TestFlight app. Filed as a signal of where the production real-time AR-video frontier is heading, not as a technical reference.

Key claims

Note: This is a product announcement, not a paper. “Key claims” below are the company’s marketing claims, not validated technical results. No numbers are released; no external evaluation exists.

• X1 is described as a real-time interactive video model where the phone camera + touchscreen replace text prompts — the camera acts as the input device and digital characters/scenes respond to natural hand/body gestures [tweet body + product description].
• Underlying architecture is described as a frame-level autoregressive DiT with an end-to-end streaming re-rendering pipeline; sampling is sped up via multi-stage distillation compression and adversarial training, with company press claiming a ~100× speedup over the (unspecified) baseline diffusion sampler [Sohu writeup of company materials].
• A unified interaction model is claimed to fuse 3D spatial relations with 2D screen-space operations so the model can parse complex touch gestures (pinch, drag, hold) as scene actions [Sohu writeup].
• Targets consumer-grade GPUs for inference at the same real-time speeds the company compares to Decart AI’s professional-GPU pipeline, with the company claiming ~10× lower inference cost [Sohu comparison writeup; not independently verified].
• Data scarcity for virtual–real fused interaction is addressed via a self-built semi-automated synthetic data pipeline for paired (real-scene, virtualized-overlay, gesture-response) clips [Sohu writeup].
• The product surface is built around four interaction modes — character placement (“次元互动” / character-on-camera), real-time style transfer of the camera feed (“世界滤镜”), photo-to-life animation, and a fourth gameplay-style mode — all driven by the same model [Sohu writeup].
• Distribution is a private beta via X-cam on iOS TestFlight, gated behind a Discord invite system [tweet body + Sohu writeup].
• Team is positioned as ex-Huawei “Genius Youth” recruits, Tsinghua KEG / HCI lab alumni, and HKUST(GZ) — i.e. a Chinese academic-spinout product team rather than a frontier-lab project [Sohu team writeup].

Method

No mechanical detail is given in the announcement itself. What the company has disclosed in press coverage is structurally close to the streaming-AR-video lineage the wiki already tracks (Self-Forcing / CausVid / Context Forcing / FlowAct-R1): a frame-level autoregressive DiT (chunkwise rather than full-window denoising), with the multi-step diffusion sampler distilled down to a small number of NFEs via a multi-stage pipeline that mixes distillation and adversarial training. The two interaction-specific pieces — a unified 3D-spatial + 2D-screen-gesture conditioning channel, and a synthetic data pipeline for virtual–real overlay clips — are described as the load-bearing additions on top of the AR-DiT backbone, but no architecture diagram, parameter count, training compute, or evaluation has been released.

The system runs against the live phone camera, which means the model must (a) ingest a streaming RGB feed, (b) fuse it with a user-uploaded asset (character image, style image), (c) condition on touchscreen gesture events, and (d) emit a re-rendered video stream at interactive frame rates — all on the device or on a server with mobile-network round-trip latency. Whether inference is on-device or remote is not disclosed.

Results

None publicly disclosed. No benchmark numbers, no ablations, no comparisons against Decart AI / Genie 3 / Self-Forcing-family backbones beyond the company’s own marketing claim of “1/10th the inference cost of Decart AI at matched real-time speed.” Demo videos exist but have not been independently reproduced. The only available evaluation signal is qualitative — the 178k-view demo tweet and the X-cam beta app.

Why it’s interesting

X1 is the first filed datapoint of a closed-source consumer-product in the real-time interactive video model space — the same regime as Genie 3 / Project Genie (Project Genie: Experimenting with infinite, interactive worlds) and Waymo World Model (The Waymo World Model: A New Frontier For Autonomous Driving Simulation), but pitched at end-user AR rather than at world simulation or autonomous driving. The technical claims (frame-level AR DiT + multi-stage distillation + structured interaction conditioning) line up structurally with the open Wan-based lineage filed under Autoregressive Video Generation (Context Forcing, PFP, FlowAct-R1, Causal Forcing/++), suggesting the open and closed sides of this frontier are converging on similar architectural skeletons but diverging sharply on what gets disclosed. The interesting new axis X1 adds — and Genie 3 / FlowAct-R1 / LingBot-World do not — is camera-feed-as-primary-input + on-device gesture conditioning, i.e. AR rather than world-simulation as the framing. File this as a market-signal entry, not a technical reference.

See also

• Autoregressive Video Generation — X1’s architecture description (frame-level AR DiT + multi-stage distillation) sits squarely in this lineage; no numbers to cite yet.
• World Foundation Models — closed-source flagship in the same regime as Genie 3 / Waymo World Model, but pitched at consumer AR rather than simulation.
• 4D Scene Generation — virtual-character-on-real-scene with gesture-driven physics response is a 4D-control problem; X1 is the product-layer counterpart to VerseCrafter / NeoVerse / SpaceTimePilot.
• Project Genie: Experimenting with infinite, interactive worlds — closest analogue on the open-Western side: also closed, also real-time interactive, also no architecture disclosure; differs in being world-simulation rather than camera-AR.
• The Waymo World Model: A New Frontier For Autonomous Driving Simulation — another closed flagship interactive video model, post-trained from Genie 3 into an AV simulator.
• FlowAct-R1: Towards Interactive Humanoid Video Generation — open-source technical counterpart to the kind of system X1 appears to be: a productized real-time AR-DiT with multi-stage distillation, structured memory, and interactive conditioning.
• Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation — the open-side theoretical justification for the kind of distillation pipeline X1’s marketing claims it built.