WorldStream Builds an Endless 3D World and Remembers Every Player Edit

The authors of WorldStream have built a 3D-world model that handles three tasks usually solved separately. It extends playable territory beyond the borders of any predefined map, executes free-form language commands in a world that is already running, and returns visited places with all the edits made to them. Planning runs on two levels, Macro and Micro, and the geometry itself is built not by a neural network but by deterministic code. The work appeared in Hugging Face Daily Papers.

Two planners and a compiler instead of head-on generation

WorldStream does not draw the world in pixels and does not output a finished mesh. The model produces a structured decision according to a typed schema, and everything else is assembled from that decision. This shapes the architecture.

Macro works at the level of a group of tiles and fixes the big things: where settlements stand, what the region's ecology is, how transport runs, and how the hydrology is laid out. This is where the system decides that a river will not break off at the seam between two pieces of the map, and that a road will lead somewhere. Micro takes this contract and specializes it down to the semantics of an individual tile. It takes into account the measured terrain and the commitments that neighboring tiles have already made.

Then local deterministic code takes over. It compiles the accepted plans into terrain, roads, settlements, vegetation, and entities. This setup makes results reproducible: with the same plans and seed, you get the same output.

Streaming with different horizons

The main engineering idea is that different stages have different ranges. Plans run far ahead along the predicted corridor of the player's movement. Compilation starts only where movement will soon require geometry. Everything left behind the user is freed from memory.

As a result, the world grows within limited budgets. The cheap level (the plan) looks far ahead, while the expensive one (geometry) stays close to the player. For an open world this is ordinary streaming logic, just applied to a generative model.

One world record instead of hidden memory

The most practical part of the work is the single world record. It holds plans, seeds, object identities and their mutable state, boundary commitments between tiles, and edits tracked by attribute. The record updates as the player moves forward and changes things. When the player returns, tiles are recompiled from the record rather than reconstructed from the model's vague "memory."

In other words, the system stores the recipe of the world and the deltas on top of it, not the finished geometry. That is where the persistence comes from: a felled tree or a built road will not vanish when the player comes back.

The same record serves as the basis for interaction. The model can act on existing entities, including NPCs whose state updates as they act. It can edit the current region or bind an intent to regions that have not been planned yet.

System components

• Planner: two levels, Macro (group of tiles) and Micro (individual tile), built on a typed decision model • Macro fixes: settlements, ecology, transport, hydrology • Micro accounts for: measured terrain and accepted commitments from neighbors • Assembly: deterministic local code (terrain, roads, settlements, vegetation, entities) • Streaming: plans ahead along the corridor, compilation on demand, geometry behind the player is freed • State: a single record with plans, seeds, object identities, and attribute-level edits

Where it could be useful and what to check

The first candidates for adoption are game studios and prototyping tools. A world that extends itself around the player, understands text commands, and keeps edits fits sandboxes, procedural RPGs, and level editors. The second scenario is simulation environments for training agents: a structured world record is easier to replay and debug than the hidden state of a video model.

Three things deserve attention in evaluation. The first is consistency at the borders of independently planned tiles over long sessions. The second is how the world record grows if the player changes the environment heavily. The third is how the planner behaves on unusual commands, when the intent is bound to a region that has not been planned yet. That is where it will become clear whether the hierarchy keeps its promise to "expand under a limited budget" without losing coherence.