DreamForge-World 0.1 Preview

What is a real-time world model?

A world model is a neural network that learns the dynamics of an environment well enough to generate what happens next. DreamForge runs this loop in real time: you provide input — move forward, turn, look around — and the model generates the next moments of the world directly, frame by frame. Nothing in the scene is a pre-built mesh, sprite, or level. The world itself is the model's output.

Our approach

DF-World 0.1 is a 1.5B-parameter model built on open research: LongLive 1 (autoregressive finetune of Wan2.1-T2V-1.3B DiT) as the video backbone, Matrix-Game 2’s residual action module adapted for first- and third-person keyboard and mouse control, a custom causal runtime on top, zero-shot multimodal conditioning, and heavy optimization for local inference on consumer GPUs.

How the model works

Multimodal input initializes a world; live keyboard and mouse control — first- or third-person — drives real-time frame generation. Reprompting can change the world mid-stream.

Prototype performance & setup

Observed ranges for the current prototype — they will change as the model and serving stack evolve. DreamForge-World 0.1 generates at 480 × 832 (480p). End-to-end preview throughput includes the full runtime path: diffusion transformer execution, action conditioning, VAE decoding, and streaming overhead.

On a single RTX 4090, the default Wan2.1 VAE path reaches roughly 8–10 FPS in bf16 (~8 GB VRAM) or ~12 FPS in fp8 (~4 GB). With the optimized LightTAEW decode path, fp8 reaches up to 14–15 FPS at about 3.5 GB VRAM. On one H100, bf16 reaches roughly 17–18 FPS; fp8 on the LightTAEW path up to about 19 FPS. The model runs fully locally — including on RTX 2060-class laptop GPUs with the quantized build. Rollouts extend beyond a minute of continuous generation. Spatial world memory is not present yet.

Comparison with recent interactive world models

How DreamForge-World 0.1 compares on core capabilities against other publicly discussed interactive world model projects.

DF-World 0.1 column highlights DreamForge-World 0.1 Preview against publicly discussed interactive world model projects. Checkmarks reflect reported support from public system descriptions and do not imply equal quality, scale, or benchmark protocol.

Cost-efficient adaptation

DF-World 0.1 Preview uses 64 hours of curated gameplay and control video across the backbone LoRA and two action-module LoRA stages — the filtered adaptation set for the preview, not the raw size of upstream corpora. Public reports rarely expose a uniform training-cost metric for interactive world models, so we compare reported training-video scale where primary sources publish hour-level figures.

By adapting open video backbones with targeted LoRA runs rather than training from scratch, DreamForge targets broad interactive capability coverage per unit of direct adaptation compute — a complementary axis to frontier-scale simulators that invest heavily in training data and persistent memory.

Reported training-video scale for systems with public hour-level training reports. Bars show raw hours and scale relative to DF-World’s 64-hour curated adaptation set. The visual scale is softly compressed above 1,000 hours while preserving ordering and separation. Matrix-Game 2.0 reports about 1,200 hours of interaction-annotated data; Genie reports a 30,000-hour filtered platformer training set.

Where it stands today

DreamForge-World 0.1 is an early but working real-time world model — a controllable, fully generated environment you can step inside and drive. Here is what the current preview delivers and what we are still building toward.

How this differs from AI game-assembly tools

Many "AI game" products procedurally assemble games from pre-made assets with agentic helpers — effectively orchestrating existing pieces into a playable layout, often limited to a single camera or genre. DreamForge takes the opposite path: the model generates the world itself — open-ended, multimodal, and controllable in real time. That is the generative path to living, interactive worlds.

What's next

We are building DF-World 0.5 now and plan to open-source it once licensing, safety, and engineering constraints allow. Near term: control that feels more natural, sound and richer detail so worlds feel alive, streaming image and video updates during rollout, synchronized audiovisual generation, worlds that respond more intelligently to what you do, and spatial memory so places stay consistent as you revisit them.

@article{ayupov2026dreamforgeworld01,
  title={DreamForge-World 0.1 Preview: A Low-Compute Real-Time Controllable World Model},
  author={Ayupov, Daniyel and Markov-Tsoy, Artur},
  journal={arXiv preprint arXiv:2606.30292},
  year={2026}
}