T2V
Generate from text
Every latent frame starts from Gaussian noise and shares one timestep. The prompt fixes the scene, objects and style; optional camera and action trajectories decide how the viewpoint and the controlled subject evolve.
Technical Report · 2026 · Open weights
Astronex‑World 1.0
Real-Time Interactive World Model Foundation
1Astronex Robotics · 2Nanjing University of Information Science and Technology
- 5Bcontrollable video DiT
- 24 fps832×480 and 1280×704
- Camera · Action · Eventcontrol interfaces
- 2×L20full five-stage training
- 1×L20real-time inference
01Overview
One 5B backbone, two model forms, three kinds of control
Abstract
We present Astronex-World 1.0, an open controllable video world-model foundation. Given a text prompt (text-to-video) or an initial observation (image-to-video), the model predicts future visual states under frame-aligned camera trajectories, continuous actions and an embodiment identifier, and it accepts text events inserted at a specified position of a rollout. The family provides a bidirectional model for full-context generation and supervision and a causal model with block-causal attention and cross-block KV caching for persistent generation. Both inherit the Wan2.2-TI2V-5B visual prior. PRoPE injects camera intrinsics and extrinsics, while a 64-dimensional action stream modulates every Transformer layer. A five-stage training path develops bidirectional camera and action control, converts the backbone to block-causal generation, transfers the dense UniPC trajectory to a few-step student, restores mixed-domain dynamics, and completes asymmetric DMD/DMD2 distribution matching. The causal model generates 832×480 video at 24 fps with eight-step UniPC, an eight-latent-frame block, local history and persistent sink frames. All five training stages run on two NVIDIA L20 48 GB GPUs, and the causal model streams in real time on one. It scores 73.5 on WBench Navi and 70.0 on WBench Full. On Full, this 5B model is above the 13.6B LongCat-Video and the 14B Helios, within one point of the 22B LTX-2.3, and above YUME 1.5, which is post-trained from the same 5B Wan2.2 prior on NVIDIA A100 GPUs. The reserved action input and output interfaces allow post-training for embodied intelligence and autonomous driving.
-
Text-to-video and image-to-video
One set of weights: T2V starts every latent frame from noise, I2V keeps an encoded reference frame clean.
-
Camera control through PRoPE
Complete intrinsics and extrinsics enter the attention of all 30 Transformer layers.
-
64-D actions and embodiment IDs
A continuous action stream modulates every layer; the reserved action input and output support post-training for embodied intelligence and autonomous driving.
-
Timed event prompts
Insert a text event at a chosen frame of a running rollout for event-driven scenario simulation.
-
Real-time causal streaming
Block-causal attention, KV caching and 8-step sampling (4-step supported) stream video on a single NVIDIA L20.
-
Open weights and recipes
Inference code, post-training recipes and weights are released under Apache 2.0.
02Demos
Generated video, straight from the released checkpoints
Cinematic clips are image-to-video generations with an explicit camera trajectory; the action-controlled group is a causal rollout where the model keeps consuming control input. Click any tile to open it at full size.
Cinematic generations camera-trajectory conditioned
Rain-soaked megacity pursuit
An unbranded electric hypercar accelerates through a near-future city at sunrise; wet asphalt, spray and reflections stay physically coherent.
Storm lighthouse at dusk
A slow push toward a solitary stone lighthouse while ocean waves break against black cliffs and the beam sweeps through rain and spray.
Portrait in a gentle breeze
A cinematic close portrait: the subject holds affectionate eye contact while a continuous breeze moves long hair in layered strands without breaking facial identity.
Alpine lake at sunset
The camera glides over wildflowers toward a turquoise lake; ripples, swaying grass and water reflections stay geometrically coherent as it advances.
Ocean cliff at golden hour
A slow slide along a colossal ocean cliff: Atlantic waves explode against dark sea arches while the horizon and rock geometry remain locked.
Open-world coastal pursuit
A red supercar holds its lane between two lane markings while three police cars and a helicopter keep their identities and spacing across the shot.
Wormhole approach
A spacecraft closes on a luminous traversable wormhole: the ring stays centered and round while blue-gold lensing rotates slowly clockwise.
Supermassive black hole
A steady approach toward a supermassive black hole; the event horizon stays circular as the accretion disk brightens and gravitational lensing intensifies.
Action-controlled rollouts causal model, key inputs shown in-frame
Coastal highway
Real-road coastal drive under continuous steering input; lane markings, barriers and daylight stay consistent through the turn.
Suburban road
A forward-only rollout on a country road, showing how the world model extends a scene while the ego motion is commanded.
Night road
Low-light driving with headlight pools, lane markings and distant buildings held across the clip — a stress test for photometric stability.
Mountain road
Vehicle-view rollouts on a mountain road, where geometry and parallax have to stay coherent under combined forward and lateral control.
Snow ridge path
A bright, low-contrast environment where structure and terrain detail are easily repainted; the causal model keeps the path readable.
Open-world game
First-person game-style footage driven by CrossFPS-style stick channels, the same control space used for WBench keyboard commands.
Character close-up
Starting from a single state, the model keeps identity, lighting and scene consistent while expression and speech motion evolve. Open the tile to hear it.
33-second continuous rollout
One uninterrupted causal rollout that keeps accepting actions and extending the future without clip stitching — the long-horizon behaviour that persistent sinks and the local window are built for.
03Method
World-model controls on top of a video DiT
Astronex-World keeps the Wan2.2-TI2V-5B latent representation and its 30-layer DiT backbone rather than repeating large-scale video pretraining, and adds world-model controls through low-rank adaptation and newly initialized modules.
Text-to-video and image-to-video
I2V
Continue from an observation
The reference frame is VAE-encoded into the first latent frame, kept clean at timestep 0 and excluded from the loss. Future frames attend to it, so appearance and layout come from the image while text, camera and action decide what happens next.
Block-causal generation
The bidirectional model attends over the complete clip. The causal model uses a block lower-triangular mask: positions inside the current block interact bidirectionally, cross-block attention sees generated history only, and completed blocks are cached as keys and values instead of being recomputed.
Camera and action conditioning
For each latent frame, PRoPE turns the camera intrinsics and extrinsics into a projective positional transform inside attention. The 64-D action and embodiment ID are projected into per-frame modulation of every layer, zero-initialized so that insertion preserves the base model.
Long-horizon context
The causal model attends to a local window of the 20 most recent latent frames. The earliest four latent frames form a persistent attention sink outside this window and are never evicted. A frustum-overlap retrieval option recalls matching history when the camera revisits a region.
Released configuration
| Model | Astronex-World-5B · bidirectional and causal forms from the same weights |
|---|---|
| Base model | Wan2.2-TI2V-5B · 30 DiT layers · 48-channel latents |
| Conditions | Text, first frame, camera intrinsics/extrinsics (PRoPE), 64-D actions with 32 embodiment IDs, event text |
| Action output | Open post-training head, 64-D per latent frame |
| Causal context | 20-frame local window plus four persistent sink frames |
| Sampling | UniPC, 8 steps by default (best quality), 4 steps supported; CFG 3.0, eight latent frames per block |
| Output | 832×480 and 1280×704 at 24 fps |
| Hardware | Trained on 2× NVIDIA L20 48 GB · real-time inference on 1× L20 |
04Event prompts
Insert an event into a running world
Camera and action streams describe how the viewpoint and the agent move; events describe what happens in the scene. The causal model accepts an event prompt at a chosen latent frame of an ongoing rollout, and the event unfolds inside the scene that already exists.
-
The past stays fixed
Frames before the switch are already generated and cached; their keys and values are not recomputed, so the established scene does not change.
-
The future reads the event
At the switch the text is re-encoded and cross-attention is rebuilt. Later blocks read the new text and still attend to the cached history, which ties the event to the existing layout and subjects.
-
Appended, not substituted
The original prompt keeps describing the scene and the added text describes only the change. Replacing the prompt would ask for a different scene instead of an event inside the current one.
1.8/255mean RGB change before the switch
9.5/255mean RGB change after the switch
Measured against an otherwise identical event-free rollout, the change stays at the requested position. Because the prefix does not depend on the event, the same history can be continued with different events or with none, and the futures can be compared — for example a sudden obstacle appearing or an object being picked up.
05Training
Five stages from a bidirectional video model to a causal few-step world model
Every stage runs on two NVIDIA L20 48 GB GPUs. Stage I produces the bidirectional release; Stages II–V turn it into the streaming causal release.
-
Stage I
Bidirectional control adaptation
PRoPE camera control and 64-D action conditioning are learned under full temporal attention.
→ bidirectional release
-
Stage II
Block-causal conversion
Temporal attention becomes block-causal and is trained with teacher forcing on clean history.
→ AR diffusion model
-
Stage III
Trajectory distillation
A frozen dense UniPC teacher is distilled into a few-step student that starts from the same noise, targeting 4-step generation.
→ few-step causal model
-
Stage IV
Mixed-domain causal SFT
Camera, game, driving and robot data restore subject motion and train both I2V and T2V modes.
→ SFT checkpoint
-
Stage V
Asymmetric DMD
Distribution matching with a bidirectional real-score teacher and a motion-preservation term.
→ causal release
Training data
| Dataset | Supervision | Role | Stages |
|---|---|---|---|
| Control2V / minWM-data | Video, camera intrinsics and extrinsics | Camera trajectories and projective geometry | I–V |
| DROID | Real robot trajectories and actions | Action and embodiment conditioning | I |
| CrossFPS | First-person video with aligned controls | Action-conditioned navigation and game motion | IV, V |
| DrivingDojo | Driving video and road scenarios | Ego motion, traffic events and road dynamics | IV, V |
| NVIDIA PhysicalAI | Synthetic embodied robot scenes | Contact, manipulation and embodied motion | IV, V |
06Results
WBench and VBench
Scores on two public benchmarks, and how a 5B model trained on two L20 GPUs compares with open models from 4B to 22B.
WBench scores
70.0Full 289 average
73.5Navi 158 average
78.3Quality
73.8Setting
47.6Interaction
82.4Consistency
68.1Physical
Dimension scores on Full 289, evaluated with the official WBench code.
Model scale and training compute WBench Full 289, live leaderboard snapshot of September 13, 2026
2× NVIDIA L20 48 GB
All five training stages of Astronex-World 1.0 run on two L20 GPUs, and the released causal model streams
832×480 video at 24 fps in real time on one.
| Model | Params | Training GPUs | Avg. | Quality | Setting | Interaction | Consistency | Physical |
|---|---|---|---|---|---|---|---|---|
| Kairos 3.0 | 4B | n/r | 65.7 | 73.1 | 70.3 | 41.6 | 83.2 | 60.5 |
| YUME 1.5† | 5B | NVIDIA A100 | 68.9 | 77.6 | 72.4 | 48.4 | 80.9 | 65.4 |
| Astronex-World 1.0† | 5B | 2× L20 48 GB | 70.0 | 78.3 | 73.8 | 47.6 | 82.4 | 68.1 |
| HY-Video 1.5 | 8.3B | n/r | 74.3 | 76.6 | 85.6 | 54.7 | 87.5 | 67.1 |
| LongCat-Video | 13.6B | n/r | 69.9 | 77.2 | 72.3 | 45.1 | 86.6 | 68.4 |
| Helios (distilled)† | 14B | 64–128× H100 | 69.7 | 73.3 | 75.3 | 41.6 | 82.2 | 76.1 |
| LTX-2.3 | 22B | n/r | 70.9 | 77.0 | 85.2 | 49.4 | 78.0 | 65.1 |
Training hardware follows each model's report (n/r: not reported). †: post-trained from a Wan prior.
Same 5B class
YUME 1.5 is post-trained from the same Wan2.2 5B prior on A100 GPUs. Astronex-World 1.0, trained on two L20 GPUs, is higher in Average (70.0 vs. 68.9) and in four of five dimensions, and 4.3 points above the 4B Kairos 3.0.
Larger models
With a third or less of the parameters, it is above LongCat-Video (13.6B) and Helios (14B) and within 0.9 points of LTX-2.3 (22B). Helios uses 64–128 H100 GPUs per stage. Quality 78.3 is the highest in the table.
What this shows
A strong video prior plus staged post-training is enough to build a real-time interactive world model that competes with models two to four times larger — reproducible on hardware most research groups have.
VBench 1.0
| Setting | Imaging Quality | Aesthetic Quality | Motion Smoothness | Temporal Flickering | Background | Camera Motion |
|---|---|---|---|---|---|---|
| Text-to-video | 0.715 | 0.508 | 0.990 | 0.987 | – | – |
| Image-to-video | 0.738 | 0.473 | – | 0.995 | 0.985 | 0.485 |
Official VBench 1.0 code; 240 T2V and 118 I2V videos generated in the official prompt order.
07Limitations & roadmap
Limitations and next steps
W → right → left sequence, and a warehouse gradually darkens and loses local detail.
Attention sinks delay but do not eliminate long-term drift.
- Weak complex interaction. Event editing, subject action and perspective switching rely on a text interface that cannot yet express multiple independent timed events in one rollout.
- Remaining long-horizon drift. Strong turns, loops and low-light scenes still show color shift, darkening, structural repainting and detail loss.
- Distillation suppresses motion. DMD can prefer stable but static solutions, far below bidirectional motion amplitude.
- Limited physical fidelity. The model has no explicit physical state, depth, collision constraint or 3D scene representation.
- Action post-training. The 64-D action stream and action-output path are reserved interfaces; driving and robot control spaces are connected through post-training with domain data.
Next
Astronex-World 2.0
2.0 will include 14B and 5B scales, both with bidirectional and causal forms and unified text, image, camera, continuous-action and embodiment interfaces. The 14B model targets visual capacity, complex dynamics and long-horizon scene modelling; the 5B model keeps the low-cost real-time path.
Primary objectives are long-term color drift, structural repainting, loop inconsistency, motion decay and complex interaction. Building on the reserved action interface, post-trained models for embodied intelligence and autonomous driving will follow, with action adherence, planning and task success added to the evaluation.
- 14B and 5B scales, bidirectional + causal
- Unified camera, action, event and embodiment interfaces
- Action post-training for embodied intelligence and autonomous driving
08Citation
BibTeX
@techreport{zhou2026astronexworld,
title = {Astronex-World 1.0: Real-Time Interactive World Model Foundation},
author = {Zhou, Xin and Miao, Cong},
institution = {Astronex Robotics},
year = {2026},
type = {Technical Report}
}