Is DreamX-World open source?

Yes. DreamX-World 1.0 is released by Alibaba's AMAP research team under the Apache-2.0 license. The team published the code on GitHub, the paper on arXiv, a project page, and two model checkpoints (5B and 5B-Cam) on Hugging Face, so researchers can download and run the system directly.

What does DreamX-World do?

DreamX-World is an interactive video world model. You give DreamX-World a starting image or prompt plus camera moves and event instructions, and DreamX-World generates a coherent video you can navigate inside. It supports free camera control, scene revisit with memory, and named event control across photoreal, game-style, and stylized worlds.

How fast is DreamX-World?

DreamX-World reaches up to 16 FPS when running on eight RTX 5090 GPUs. The model uses optimizations such as mixed-precision DiT execution, residual reuse, 75%-pruned VAE decoding, and asynchronous pipeline parallelism. A technique the authors call Efficient PRoPE cuts inference latency by about 30 percent compared with full PRoPE.

How does DreamX-World compare to other world models?

On its evaluation, DreamX-World scores 84.76 overall, ahead of HY-WorldPlay 1.5 at 80.79 and LingBot-World at 80.45, with a camera-control score of 73.75. AMAP reports a 57 to 62 percent human-preference win rate over HY-WorldPlay 1.5 and LingBot-World on the same evaluation. Earlier comparable systems, such as Project Genie and Decart Oasis, have mostly stayed closed.

AMAP-ML · 2026-06-15 · notable

DreamX-World 1.0 — Alibaba AMAP open-sources an interactive world model

Alibaba's AMAP research team releases a 5B Apache-2.0 video world model with camera navigation, scene revisit, and event control across photoreal, game-style, and stylized domains. Code, paper, and two checkpoints are out.

DreamX-World repository banner from AMAP-ML on GitHub

Open-source 5B world model that lets you steer the camera, revisit a scene, and stage events across photoreal, game, and stylized worlds.

Key specs

Parameters	5B
GitHub stars	325
Camera control score	73.75
Overall score	84.76
Fps 8x rtx5090	16

Quick facts

Maker	AMAP-ML (Alibaba AMAP research team)
License	Apache-2.0 (open source)
Parameters	5B
Checkpoints	Two (5B and 5B-Cam) on Hugging Face
Speed	Up to 16 FPS on eight RTX 5090 GPUs
Overall score	84.76 (vs 80.79 HY-WorldPlay 1.5, 80.45 LingBot-World)
What it does	Interactive video world model with camera control, scene revisit, and event control

What is it?

DreamX-World 1.0 is an interactive video world model from Alibaba's AMAP research group. You give it a starting image or prompt plus camera moves and event instructions, and it generates a coherent video the agent or player can navigate inside.

How does it work?

The model uses a progressive training pipeline that adds camera-aware conditioning, geometry-guided memory for scene revisit, structured event instruction tuning, autoregressive long-video generation with distillation, and a reinforcement-learning quality pass. A technique the authors call Efficient PRoPE projects positional encodings to spatially reduced tokens, cutting inference latency by about 30% versus full PRoPE.

Why does it matter?

World models that combine free camera control with memory across revisits and named events have mostly been closed (Project Genie, Decart Oasis). Shipping a 5B Apache-2.0 checkpoint that reaches 16 fps on eight RTX 5090s puts a comparable system in researchers' hands. AMAP reports a 57–62% human-preference win rate over HY-WorldPlay 1.5 and LingBot-World on the same evaluation.

Who is it for?

world-model researchers, robotics simulation teams, game engine experiments

Frequently asked questions

Is DreamX-World open source?: Yes. DreamX-World 1.0 is released by Alibaba's AMAP research team under the Apache-2.0 license. The team published the code on GitHub, the paper on arXiv, a project page, and two model checkpoints (5B and 5B-Cam) on Hugging Face, so researchers can download and run the system directly.
What does DreamX-World do?: DreamX-World is an interactive video world model. You give DreamX-World a starting image or prompt plus camera moves and event instructions, and DreamX-World generates a coherent video you can navigate inside. It supports free camera control, scene revisit with memory, and named event control across photoreal, game-style, and stylized worlds.
How fast is DreamX-World?: DreamX-World reaches up to 16 FPS when running on eight RTX 5090 GPUs. The model uses optimizations such as mixed-precision DiT execution, residual reuse, 75%-pruned VAE decoding, and asynchronous pipeline parallelism. A technique the authors call Efficient PRoPE cuts inference latency by about 30 percent compared with full PRoPE.
How does DreamX-World compare to other world models?: On its evaluation, DreamX-World scores 84.76 overall, ahead of HY-WorldPlay 1.5 at 80.79 and LingBot-World at 80.45, with a camera-control score of 73.75. AMAP reports a 57 to 62 percent human-preference win rate over HY-WorldPlay 1.5 and LingBot-World on the same evaluation. Earlier comparable systems, such as Project Genie and Decart Oasis, have mostly stayed closed.

Try it

https://github.com/AMAP-ML/DreamX-World