| title | description |
|---|---|
Instance streaming |
Instance streaming allows the Roblox Engine to dynamically load and unload 3D content in regions of the world. |
In-experience instance streaming allows the Roblox Engine to dynamically load and unload 3D content and related instances in regions of the world. This can improve the overall player experience in several ways, for example:
- Faster join times — Players can start playing in one part of the world while more of the world loads in the background.
- Memory efficiency — Experiences can be played on devices with less memory since content is dynamically streamed in and out. More immersive and detailed worlds can be played on a wider range of devices.
- Improved performance — Better frame rates and performance, as the server can spend less time and bandwidth synchronizing changes between the world and players in it. Clients spend less time updating instances that aren't currently relevant to the player.
- Level of detail — Distant models and terrain remain visible even when they're not streamed to clients, keeping the experience optimized without entirely sacrificing background visuals.
Instance streaming is enabled through the StreamingEnabled property of the Workspace object in Studio. This property cannot be set in a script. Streaming is enabled by default for new places created in Studio.
Once enabled, it's recommended that you adhere to the following practices:
- Because clients will not typically have the entire
Class.Workspaceavailable locally, use the appropriate tool/API to ensure that instances exist before attempting to access them in aClass.LocalScript. For example, utilize per‑model streaming controls, detect instance streaming, or useClass.Instance:WaitForChild()|WaitForChild()on objects that may not exist. - Minimize placement of 3D content outside of
Class.Workspace. Content in containers such asClass.ReplicatedStorageorClass.ReplicatedFirstis ineligible for streaming and may negatively impact join time and memory usage. - If you move a player's character by setting its
Datatype.CFrame, do so from a server-sideClass.Scriptand use streaming requests to more quickly load data around the character's new location. - Manually set replication foci only in unique situations such as experiences that don't use a
Class.Player.Characteror where streaming should occur in multiple areas of the experience. In these cases, make sure the foci are near objects that the player controls or those that should continue simulating physically on the client, and try to minimize the overall number of foci used.
By default, when a player joins an experience with instance streaming enabled, instances in the Class.Workspace are replicated to the client, excluding the following:
Class.Part|PartsorClass.MeshPart|MeshParts- Atomic, Persistent, or PersistentPerPlayer models
- Descendants of the above instances
- Non-replicating instances
Then, during gameplay, the server may stream necessary instances to the client, as they are needed.
1 Terrain is treated uniquely, in that the instance replicates to the client when the experience loads, but terrain regions only stream in when neededModels set to non-default behavior like Atomic stream in under special rules as outlined in per‑model streaming controls. However, default (nonatomic) models are sent differently based on whether ModelStreamingBehavior is set to Default (Legacy) or Improved.
When ModelStreamingBehavior is set to Default/Legacy, the Class.Model container and its non‑spatial descendants such as Class.Script|Scripts replicate to the client when the player joins. Then, when eligible, the model's Class.BasePart descendants stream in.
When ModelStreamingBehavior is set to Improved, model streaming behavior varies by whether the model is spatial (contains Class.BasePart descendants) or non‑spatial (contains no Class.BasePart descendants).
-
Instead of on player join, a spatial model (one containing
Class.BasePartdescendants) is sent only when one of itsClass.BasePartdescendants is eligible to stream in. At that point, the model and part are replicated, along with the model's non‑spatial descendants. Then, when eligible, the model's other spatial descendants stream in. -
A key consideration is when a spatial model and all of its
Class.BasePart|BasePartsbelong to a single network ownership unit, such as an avatar or NPC character model. In such cases, the entire model will stream in atomically. -
For a non‑spatial model (one without
Class.BasePartdescendants), the model container and its descendants are replicated to the client soon after the player joins, and all are exempt from streaming out. Assuming the model exists inClass.Workspacewhen the player joins, this occurs before theClass.Workspace.PersistentLoadedevent fires.
During gameplay, a client may stream out (remove from the player's Class.Workspace) regions and the Class.BasePart|BaseParts contained within them, based on the behavior set by StreamOutBehavior. The process begins with regions furthest away from the replication foci and moves in closer as needed. Regions inside the StreamingMinRadius range never stream out.
When an instance streams out, it is parented to nil so that any existing Luau state will reconnect if the instance streams back in. As a result, removal signals such as Class.Instance.ChildRemoved|ChildRemoved or Class.Instance.DescendantRemoving|DescendantRemoving fire on its parent or ancestor, but the instance itself is not destroyed in the same sense as an Class.Instance:Destroy() call.
To further anticipate stream out, examine these scenarios:
| Scenario | Example | Streaming behavior |
|---|---|---|
| A part is **created** locally through `Class.Instance.new()` in a `Class.LocalScript`. | In a "capture the flag" game, you create and attach blue helmet parts to all players on the blue team through a `Class.LocalScript`. | The part is not replicated to the server, and it is exempt from streaming out **unless** you make it a descendant of a part that exists on the server, such as a part within a player's character model. |
| A part is **cloned** locally from `Class.ReplicatedStorage` through `Class.Instance:Clone()` in a `Class.LocalScript`. | A wizard character casts a spell by activating a `Class.Tool`, upon which an object including several [special effects](../environment/index.md#special-effects) is cloned from `Class.ReplicatedStorage` and parented to the workspace at the wizard's position. | The part is not replicated to the server, and it is exempt from streaming out **unless** you make it a descendant of a part that exists on the server. |
| A part is **reparented** from `Class.ReplicatedStorage` to the workspace via a `Class.LocalScript`. | A "wizard's hat" is stored in `Class.ReplicatedStorage`. When a player chooses to play on the wizard's team, the hat is moved into their character model via a `Class.LocalScript`. | The part remains eligible for streaming out since it came from the server and was replicated to `Class.ReplicatedStorage`. Avoid this pattern as it causes a desync between the client and server, and the part may stream out; instead, **clone** the part. |
If you set ModelStreamingBehavior to Improved, the engine may stream out Default (Nonatomic) models when they're eligible to stream out, potentially freeing up memory on the client and reducing the instances which need property updates.
Under Improved model streaming behavior, streaming out of Default (Nonatomic) models is based on whether the model is spatial (contains Class.BasePart descendants) or non‑spatial (contains no Class.BasePart descendants).
- A spatial model only streams out completely when its last remaining
Class.BasePartdescendant streams out, since some of the model's spatial parts may be near to the player/replication focus and some far away. - A non‑spatial model only streams out when an ancestor streams out, equivalent to legacy streaming out behavior.
When at least one part of an assembly is eligible for streaming in, all of the assembly's parts also stream in. However, an assembly will not stream out until all of its parts are eligible for streaming out. During streaming, all of the Class.Constraint|Constraints and Class.Attachment|Attachments descending from Class.BasePart|BaseParts and atomic or persistent Class.Model|Models also stream, helping to ensure consistent physics updates on clients.
Note that assemblies with anchored parts are treated slightly differently than assemblies with only unanchored parts:
| Assembly composition | Streaming behavior |
|---|---|
| Unanchored parts only | Entire assembly is sent as an atomic unit. |
| Anchored [root part](../physics/assemblies.md#assembly-root-part) | Only the parts, attachments, and constraints needed to link the streamed parts to the root part are streamed in together. |
There may be a slight delay of ~10 milliseconds between when a part is created on the server and when it gets replicated to clients. In each of the following scenarios, you may need to use Class.Instance:WaitForChild()|WaitForChild() and other techniques rather than assuming that events and property updates always occur at the same time as part streaming.
| Scenario | Example | Streaming behavior |
|---|---|---|
| A `Class.LocalScript` makes a `Class.RemoteFunction` call to the server to create a part. | A player activates a `Class.Tool` locally to spawn a part on the server that all players can see and interact with. | When the remote function returns to the client, the part may not yet exist, even though the part is close to the client focus and within a streamed area. |
| A part is added to a character model on the server via a `Class.Script` and a `Class.RemoteEvent` is fired to a client. | When a player joins the police team, a "police badge" part stored in `Class.ServerStorage` is cloned and attached to the player's character model. A `Class.RemoteEvent` is fired and received by that player's client in order to update a local UI element. | Although the client receives the event signal, there is no guarantee that the part has already streamed to that client. |
| A part collides with an invisible region on the server and triggers a `Class.RemoteEvent` on the client. | A player kicks a soccer ball into a goal, triggering a "goal scored" event. | Other players that are close to the goal may see the "goal scored" event before the ball has been streamed to them. |
The following properties control how instance streaming applies to your experience. All of these properties are non-scriptable and must be set on the Workspace object in Studio.
Controls whether Default (Nonatomic) models are replicated when a player joins, or are only sent when needed. If this property is set to Improved, models in Class.Workspace will only be sent to clients when needed, potentially speeding up join times. See Technical Behavior for more details.
Your experience may behave in unintended ways if a player moves into a region of the world that hasn't been streamed to them. The streaming integrity feature offers a way to avoid those potentially problematic situations. Please see the Enum.StreamingIntegrityMode documentation for more details.
The StreamingMinRadius property indicates the radius around the replication foci in which instances stream in at the highest priority. Care should be taken when increasing the default, as doing so will require more memory and more server bandwidth at the expense of other components.
The StreamingTargetRadius property controls the maximum distance away from the replication foci in which instances stream in. Note that the engine is allowed to retain previously loaded instances beyond the target radius, memory permitting.
A smaller StreamingTargetRadius reduces server workload, as the server will not stream in additional instances beyond the set value. However, the target radius is also the maximum distance players will be able to see the full detail of your experience, so you should pick a value that creates a nice balance between these.
**StreamingTargetRadius** should be larger than **StreamingMinRadius**. 3D content between the target radius and the minimum radius acts as a buffer in case the client temporarily stops receiving new content from the server. When the minimum radius and the target radius are equal, there is no buffer, which can lead to an increase in network pauses or an otherwise suboptimal user experience.The StreamOutBehavior property sets the streaming out behavior according to one of the following values:
| Setting | Streaming behavior |
|---|---|
| **Default** | Default behavior, currently the same as **LowMemory**. |
| **LowMemory** | The client only streams out parts in a low memory situation and may remove 3D content until only the minimum radius is present. |
| **Opportunistic** | Regions beyond [StreamingTargetRadius](#streamingtargetradius) can be removed on the client even when there is no memory pressure. In this mode, the client never removes instances that are closer than the target radius, except in low memory situations. |
By default, streaming occurs around the local player's character's Class.Model.PrimaryPart|PrimaryPart, although you can specify a different replication focus point through Class.Player.ReplicationFocus.
You can also add and remove additional replication foci through Class.Player:AddReplicationFocus() and Class.Player:RemoveReplicationFocus() to dynamically enable streaming in multiple areas of the experience.
On the client, too many foci for a player can limit the engine's ability to adjust to memory limitations and make it more likely for clients to be killed by the OS for using too much memory.
Client-side physics simulation only occurs in streamed areas, even for locally created instances and for [persistent](#persistent) instances. If you have instances that you'd like to keep simulating even when they're far away from the character, create an additional replication focus near those instances. In many experiences, it's common for players to move back and forth between the same areas frequently, for example between their "home base" and a "trading hub." In such cases, you can create a replication focus point in each area to ensure those areas are readily present on client devices. Multiple replication points are useful when players can view specific, important regions through a scope, such as enemy bases scattered across a barren landscape. In such cases, you can create a replication focus point in each base to ensure players see details and simulated physics from afar.Globally, the ModelStreamingBehavior property lets you control how models are streamed in on join. Additionally, to avoid issues with streaming on a per-model basis and minimize use of Class.Instance:WaitForChild()|WaitForChild(), you can customize how Class.Model|Models and their descendants stream through their Class.Model.ModelStreamingMode|ModelStreamingMode property.
When a Class.Model is set to Default or Nonatomic, streaming behavior varies based on whether ModelStreamingBehavior is set to Default (Legacy) or Improved.
| [ModelStreamingBehavior](#modelstreamingbehavior) | Technical behavior |
|---|---|
| **Default** (**Legacy**) | The model is replicated when a player joins. This potentially results in more instances sent during loading, more instances stored in memory, and additional complexity for scripts that want to access the model's descendants. For example, a separate `Class.LocalScript` will need to use `Class.Instance:WaitForChild()|WaitForChild()` on a descendant `Class.BasePart` inside the model. |
| **Improved** | The model is only sent when needed, potentially speeding up join times. |
See technical behavior for more details.
If a Class.Model is changed to Atomic, all of its descendants are streamed in together when a descendant Class.BasePart is eligible. As a result, a separate Class.LocalScript that needs to access instances in the model would need to useClass.Instance:WaitForChild()|WaitForChild() on the model itself, but not on a descendant Class.MeshPart or Class.Part since they are sent alongside the model.
An atomic model is only streamed out when all of its descendant parts are eligible for streaming out, at which point the entire model streams out together. If only some parts of an atomic model would typically be streamed out, the entire model and its descendants remain on the client.
local Workspace = game:GetService("Workspace")
-- Atomic model does not exist at load time; use WaitForChild()
local model = Workspace:WaitForChild("Model")
-- Descendant parts stream in with model and are immediately accessible
local meshPart = model.MeshPart
local part = model.PartPersistent models are not subject to normal streaming in or out. They are sent as a complete atomic unit soon after the player joins and before the Class.Workspace.PersistentLoaded event fires. Persistent models and their descendants are never streamed out, but to safely handle streaming in within a separate Class.LocalScript, you should use Class.Instance:WaitForChild()|WaitForChild() on the parent model, or wait for the Class.Workspace.PersistentLoaded|PersistentLoaded event to fire.
local Workspace = game:GetService("Workspace")
-- Persistent model does not exist at load time; use WaitForChild()
local model = Workspace:WaitForChild("Model")
-- Descendant parts stream in with model and are immediately accessible
local meshPart = model.MeshPart
local part = model.PartAvoid creating catch-all persistent models that have a large number of sub-models. For example, if you're creating an experience with a large number of vehicles, do not create a single persistent model which contains every vehicle in the experience that you want to be persistent. Instead, set each individual vehicle model to be persistent, if necessary.
Runtime performance impacts of persistent models after replication are mostly the same as regular models without streaming enabled. An exception is when the contents of the model change frequently or if parts in the model have changes to their physical connections. In these cases, the engine must perform extra updates to ensure those changes are reflected correctly on all clients, so it's best to avoid frequent changes to persistent models.Models set to PersistentPerPlayer behave the same as Persistent for players that have been added using Class.Model:AddPersistentPlayer(). For other players, behavior is the same as Atomic. You can revert a model from player persistence via Class.Model:RemovePersistentPlayer().
If you set the Datatype.CFrame of a player character to a region which isn't currently loaded, streaming pause occurs, if enabled through Enum.StreamingIntegrityMode. If you know the character will be moving to a specific area, you can call Class.Player:RequestStreamAroundAsync() to request that the server sends regions around that location to the client.
The following scripts show how to fire a client-to-server remote event to teleport a player within a place, yielding at the streaming request before moving the character to a new Datatype.CFrame.
local ReplicatedStorage = game:GetService("ReplicatedStorage")
local teleportEvent = ReplicatedStorage:WaitForChild("TeleportEvent")
local function teleportPlayer(player, teleportTarget)
-- Request streaming around target location
player:RequestStreamAroundAsync(teleportTarget)
-- Teleport character
local character = player.Character
if character and character.Parent then
local currentPivot = character:GetPivot()
character:PivotTo(currentPivot * CFrame.new(teleportTarget))
end
end
-- Call teleport function when the client fires the remote event
teleportEvent.OnServerEvent:Connect(teleportPlayer)local ReplicatedStorage = game:GetService("ReplicatedStorage")
local teleportEvent = ReplicatedStorage:WaitForChild("TeleportEvent")
local teleportTarget = Vector3.new(50, 2, 120)
-- Fire the remote event
teleportEvent:FireServer(teleportTarget)If players can encounter streaming pauses in your experience, you might want to customize the pause screen. The Class.Player.GameplayPaused property indicates the player's current pause state. This property can be used with a Class.Instance:GetPropertyChangedSignal()|GetPropertyChangedSignal() connection to show or hide a custom GUI.
local Players = game:GetService("Players")
local GuiService = game:GetService("GuiService")
local player = Players.LocalPlayer
-- Disable default pause modal
GuiService:SetGameplayPausedNotificationEnabled(false)
local function onPauseStateChanged()
if player.GameplayPaused then
-- Show custom GUI
else
-- Hide custom GUI
end
end
player:GetPropertyChangedSignal("GameplayPaused"):Connect(onPauseStateChanged)In some cases, it's necessary to detect when an object streams in or out and react to that event. A useful pattern for streaming detection is as follows:
-
Using the Tags section of an instance's properties, or Studio's Tag Editor, assign a logical
Class.CollectionServicetag to all of the affected objects. -
From a single
Class.LocalScript, detect when a tagged object streams in or out throughClass.CollectionService:GetInstanceAddedSignal()|GetInstanceAddedSignal()andClass.CollectionService:GetInstanceRemovedSignal()|GetInstanceRemovedSignal(), then handle the object accordingly. For example, the following code adds taggedClass.Lightobjects into a "flicker" loop when they stream in and removes them when they stream out.local CollectionService = game:GetService("CollectionService") local tagName = "FlickerLightSource" local random = Random.new() local flickerSources = {} -- Detect currently and new tagged parts streaming in or out for _, light in CollectionService:GetTagged(tagName) do flickerSources[light] = true end CollectionService:GetInstanceAddedSignal(tagName):Connect(function(light) flickerSources[light] = true end) CollectionService:GetInstanceRemovedSignal(tagName):Connect(function(light) flickerSources[light] = nil end) -- Flicker loop while true do for light in flickerSources do light.Brightness = 8 + random:NextNumber(-0.4, 0.4) end task.wait(0.05) end
When streaming is enabled, Class.Model|Models outside of the currently streamed area will not be visible by default. However, you can instruct the engine to render lower resolution "imposter" meshes for models that are not present on clients through each model's Class.Model.LevelOfDetail|LevelOfDetail property.
| Model setting | Streaming behavior |
|---|---|
| **StreamingMesh** | Activates the asynchronous generation of an imposter mesh to display when the model is not present on clients. |
| **Disabled** / **Automatic** | The model disappears when outside the streaming radius. |
When using imposter meshes, note the following:
- Imposter meshes are designed to be seen at 1024 studs away from the camera or further. If you've reduced StreamingTargetRadius to a much smaller value like 256, imposter meshes may not be visually acceptable for the model they replace.
- If a model and its descendant models are all set to StreamingMesh, only the top-level ancestor model is rendered as an imposter mesh, wrapping all geometries under the ancestor as well as its descendant models. For better performance, it's recommended that you use Disabled for descendant models.
- Textures are not supported; imposter meshes are rendered as smooth meshes.
- While a
Class.Modelis not completely streamed in, the imposter mesh is rendered instead of individual parts of the model. Once all individual parts are streamed in, they render and the imposter mesh is ignored. - Imposter meshes have no physical significance and they act as non-existent with respect to raycasting, collision detection, and physics simulation.
- Editing a model in Studio, such as adding/deleting/repositioning child parts or resetting colors, automatically updates the representative mesh.







