DreamX-World generates explorable worlds you steer with camera and prompts

A world model for interactive simulation, not a clip generator

DreamX-World is described as a general-purpose world model for interactive world simulation. It generates worlds that a user can explore, control and transform through event prompts, which is a different goal from producing a short video clip from a text description. The distinction is the point of the project. A clip is a fixed artefact; a world is something you move through, and the interesting technical problem becomes keeping that world coherent while the viewer changes the camera and triggers events. The project is published under Apache-2.0, with the model weights distributed on Hugging Face and ModelScope under the GD-ML organisation, a technical report on arXiv as 2606.16993, and a project page and hosted experience linked from the README. The model family is named after its size, with a 5B checkpoint released as the base for long-horizon generation and a 5B-Cam variant for camera-driven generation, while a 14B-Cam model and an audio-video joint generation model are listed as still planned.

Progressive training from action control to event response

The training story in the README is unusually specific, and it explains the behaviour of the released checkpoints. The data comes from a scalable data engine working on Unreal Engine data, gameplay footage and real-world videos, combined with camera estimation and strict data filtering, so the model learns realistic dynamics and interactions rather than copying frames. The curriculum is progressive. The first stage is fine-grained action control, the second is open-ended event response, and reinforcement learning is then used to improve action following, interaction consistency and visual fidelity. Finally, forcing and distillation are applied so that inference is efficient enough for interactive generation at scale. The interesting design choice is the ordering: strict control of camera movement is learned before the model is allowed to improvise, which is the opposite of how a text-to-video model is usually trained and is what makes the long-horizon behaviour possible.

Two released checkpoints, two different generation strategies

The checkpoint table lists two models and the difference between them is the approach, not the size. DreamX-World-5B-Cam is bidirectional and supports 5 second video generation with camera control. DreamX-World-5B is autoregressive and supports long-horizon video generation. Both are available on Hugging Face and ModelScope, and both point to the same inference_README.md for detailed instructions. If you want camera moves over a short clip, the bidirectional model is the one to reach for. If you want a world that keeps going as the user keeps exploring, the autoregressive model is the one, and the README claims coherent world exploration over hundreds of frames with precise camera control. Progressive training on long rollouts is what mitigates the usual failure modes of autoregressive video, which are drift in identity, background, style and colour as the sequence lengthens.

Geometry-guided memory retrieval for revisited scenes

The second long-horizon problem is that a viewer who walks back to a place they have already seen expects it to look the same. DreamX-World addresses this with geometry-guided memory retrieval, which recovers non-local visual evidence from earlier observations. The effect described in the documentation is scene persistence when the camera revisits a previously explored region, with the layout, object identities and local appearance preserved rather than regenerated from scratch. This is a geometric rather than purely temporal mechanism: the model uses spatial information to decide which earlier frames are relevant to what it is looking at now, instead of pulling from the most recent context window. For anyone evaluating interactive world models, this and the drift mitigation are the two features worth measuring, because they are what separate a walkable scene from a sequence of unrelated shots.

Setup is a pip install plus base checkpoints

The quick start in the README is two steps. First the dependencies:

pip install -r requirements.txt

Second, the Wan2.2-5B-TI2V checkpoints, which are downloaded from the Wan-AI organisation on Hugging Face. Everything else is delegated to inference_README.md, which is where the detailed instructions live. That delegation is worth flagging, because the inference command in the top-level quick start block is commented out in the README, so following the visible text alone gets you as far as the install. The repository itself contains the entry points: inference_dreamx5b.py with inference_dreamx_5b.sh beside it, and a separate inference_ar_forcing.py with inference_ar_forcing.sh, which is the forcing and distillation path mentioned in the training description. The dependency set also fixes the environment tightly. The requirements file pins torch at 2.5.1 and torchvision at 0.20.1, and pins xfuser at 0.4.1, flash_attn at 2.8.3 and triton at 3.1.0, with diffusers, transformers, accelerate and gradio for the surrounding stack, so an existing environment will need to match.

What a generated world costs on one H20

The efficiency table gives the numbers, with the caveat that the reported cost covers both denoising and VAE decoding time. On a single H20, the bidirectional DreamX-World-5B-Cam model takes 509 seconds and 38 GB of peak memory for 5 seconds of 720P video. The autoregressive DreamX-World-5B model produces the same 5 seconds of 720P in 26 seconds with 40 GB of peak memory, and a full 60 seconds of 720P in 342 seconds with 72 GB of peak memory. The gap between the two models on the same task is dramatic, and it tracks the generation strategy rather than the parameter count. The 60 second run is the one that matters for interactive use, and it is also the one that crosses from 40 GB to 72 GB of memory. There are no GitHub releases, so the checkpoints are the artefacts to pin, and the last push to the repository was on 2026-07-23.

Editorial conclusion

Adopt DreamX-World if you are researching interactive world simulation and you have access to an H20 class GPU, because the open weights let you study camera control and long-horizon behaviour without running the live service. Do not adopt it for production video work on modest hardware, since the bidirectional model takes 509 seconds and 38 GB of memory for five seconds of 720P output. Before you start, read inference_README.md rather than the top-level quick start, check that your driver stack matches the pinned flash_attn and triton versions, and decide first between the bidirectional 5B-Cam checkpoint and the autoregressive 5B checkpoint.

Frequently asked questions

Official sources

1. AMAP-ML/DreamX-World on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README