pixijs-performance

---

name: pixijs-performance description: "Use this skill when profiling or optimizing a PixiJS v8 app for FPS, draw calls, or GPU memory. Covers destroy patterns (cacheAsTexture(false), releaseGlobalResources), GCSystem and TextureGCSystem, PrepareSystem, object pooling, batching rules, BitmapText for dynamic text, culling (Culler, CullerPlugin, cullable, cullArea), resolution/antialias tradeoffs, low-level wins (transient MSAA render textures, render bundles, partial buffer updates). Triggers on: FPS, jank, draw calls, batching, object pool, GCSystem, PrepareSystem, Culler, cacheAsTexture, memory leak, destroy patterns, render bundle, transient, MSAA, Buffer.update, WebGPU performance, TexturePool, repeatEdgePixels, bind group." license: MIT

Profile before optimizing. PixiJS handles a lot of content well out of the box; browser DevTools Performance + GPU profiling should be your first move. Once you've found the bottleneck, apply the targeted pattern below (destroy, pool, batch, cache, or cull).

Quick Start

container.cacheAsTexture(true);
container.updateCacheTexture();
container.cacheAsTexture(false);
container.destroy({ children: true });

import { CullerPlugin, extensions } from "pixi.js";
extensions.add(CullerPlugin);

offscreenContainer.cullable = true;
offscreenContainer.cullArea = new Rectangle(0, 0, 256, 256);

// Tune GC via init options (ms). The `textureGC.*` properties are
// deprecated since 8.15.0 — use these on the Application init instead.
await app.init({ gcMaxUnusedTime: 60_000, gcFrequency: 30_000 });

Related skills: pixijs-scene-container (destroy options), pixijs-scene-core-concepts (render groups, layers, culling), pixijs-scene-text (BitmapText for dynamic content), pixijs-assets (atlasing), pixijs-custom-rendering (custom batchers).

Core Patterns

Proper destroy with cleanup

import { Sprite, Assets } from "pixi.js";

const texture = await Assets.load("character.png");
const sprite = new Sprite(texture);

// Destroy sprite only (preserve texture for reuse)
sprite.destroy();

// Destroy sprite AND its texture
sprite.destroy({ children: true, texture: true, textureSource: true });

When done with a loaded asset entirely:

Assets.unload("character.png");

This removes it from the cache and unloads the GPU resource.

Objects you build for custom rendering clean up the same way: call geometry.destroy() and renderTarget.destroy() on the ones you create. Destroying a container also destroys the batchers cached for its render group.

Application destroy/recreate cycle

import { Application } from "pixi.js";

// Correct destroy that cleans global pools
app.destroy({ releaseGlobalResources: true });

const newApp = new Application();
await newApp.init({ width: 800, height: 600 });

Without releaseGlobalResources: true, pooled objects (batches, textures) from the old app leak into the new one, causing flickering and corruption.

Texture garbage collection

PixiJS auto-collects unused textures and GPU resources (including WebGPU bind groups) via GCSystem. Defaults: checks every 30 seconds, removes resources idle for 60 seconds. These are time-based (milliseconds).

import { Application } from "pixi.js";

const app = new Application();

await app.init({
  gcActive: true,
  gcMaxUnusedTime: 120000, // idle time before cleanup in ms (default: 60000)
  gcFrequency: 60000, // check interval in ms (default: 30000)
});

For manual control:

texture.source.unload(); // immediate GPU memory release

PrepareSystem for GPU upload

Upload textures and graphics to GPU before rendering to avoid first-frame hitches:

import "pixi.js/prepare";
import { Application, Assets } from "pixi.js";

const app = new Application();
await app.init();

// Don't render until assets are uploaded
app.stop();

const texture = await Assets.load("large-scene.png");

// Upload to GPU ahead of time
await app.renderer.prepare.upload(app.stage);

// Now rendering won't hitch on first frame
app.start();

prepare.upload() accepts a Container (uploads all textures, text, and graphics in the subtree) or individual resources.

cacheAsTexture for performance

cacheAsTexture() renders a container's subtree to a single texture, reducing draw calls for complex static content. Internally it creates a render group and caches the result.

When to use:

• Many static children (UI panels, decorative backgrounds, complex Graphics)
• Containers with expensive filters (cache the filter result)
• Large subtrees that rarely change

Tradeoffs:

• Uses GPU memory for the cached texture (larger containers = more memory)
• Max texture size is GPU-dependent (typically 4096x4096 or larger). No PixiJS property exposes it, so query the backend directly:
• Must call updateCacheTexture() after modifying children
• Combining with masks is fragile (see the masking skill)

import type { WebGLRenderer, WebGPURenderer } from "pixi.js";

// Available after renderer/app init on both backends
const maxTextureSize = renderer.name === "webgpu"
  ? (renderer as WebGPURenderer).gpu.device.limits.maxTextureDimension2D
  : (renderer as WebGLRenderer).gl.getParameter(WebGL2RenderingContext.MAX_TEXTURE_SIZE);

import { Container, Sprite } from "pixi.js";

const panel = new Container();
// ... add many static children ...

panel.cacheAsTexture(true);

// With options
panel.cacheAsTexture({ resolution: 2, antialias: true });

// Refresh after changes
panel.updateCacheTexture();

// MUST disable before destroying (see Common Mistakes below)
panel.cacheAsTexture(false);
panel.destroy();

Avoid: toggling on/off repeatedly (constant re-caching negates benefits), caching sparse containers (negligible gain), caching containers larger than 4096x4096.

Object recycling

Reuse objects by changing their properties instead of destroy/recreate:

import { Sprite, Container, Texture } from "pixi.js";

class BulletPool {
  private _pool: Sprite[] = [];
  private _container: Container;

  constructor(container: Container) {
    this._container = container;
  }

  public get(texture: Texture): Sprite {
    let bullet = this._pool.pop();

    if (!bullet) {
      bullet = new Sprite(texture);
      this._container.addChild(bullet);
    }

    bullet.texture = texture;
    bullet.position.set(0, 0);
    bullet.rotation = 0;
    bullet.scale.set(1);
    bullet.alpha = 1;
    bullet.tint = 0xffffff;
    bullet.blendMode = "normal";
    bullet.visible = true;
    return bullet;
  }

  public release(bullet: Sprite): void {
    bullet.visible = false;
    this._pool.push(bullet);
  }
}

Destroying and recreating is significantly more expensive than toggling visible and updating properties. GPU resources stay allocated; only scene graph visibility changes.

Batching optimization

PixiJS batches similar consecutive objects into single draw calls. Batch breaks occur on:

• Object type change (Sprite vs Graphics)
• Texture source change (beyond the per-batch texture limit, typically 16)
• Blend mode change
• Topology change

Optimize draw order:

import { Sprite, Graphics, Container } from "pixi.js";

// 4 draw calls: type alternates
const bad = new Container();
bad.addChild(new Sprite(t1));
bad.addChild(new Graphics().rect(0, 0, 10, 10).fill(0xff0000));
bad.addChild(new Sprite(t2));
bad.addChild(new Graphics().rect(0, 0, 10, 10).fill(0x00ff00));

// 2 draw calls: types grouped
const good = new Container();
good.addChild(new Sprite(t1));
good.addChild(new Sprite(t2));
good.addChild(new Graphics().rect(0, 0, 10, 10).fill(0xff0000));
good.addChild(new Graphics().rect(0, 0, 10, 10).fill(0x00ff00));

Same principle applies to blend modes: screen/normal/screen/normal = 4 draws; screen/screen/normal/normal = 2 draws.

Spritesheets over individual textures

import { Assets, Sprite } from "pixi.js";

// Load a spritesheet (single texture atlas)
const sheet = await Assets.load("game-atlas.json");

// All frames share one GPU texture; enables batching
const hero = new Sprite(sheet.textures["hero.png"]);
const enemy = new Sprite(sheet.textures["enemy.png"]);
const coin = new Sprite(sheet.textures["coin.png"]);

Individual textures each require their own GPU upload and break batches when the texture limit is exceeded. Spritesheets consolidate many frames into one atlas texture.

Use @0.5x filename suffix on half-resolution sheets so PixiJS auto-scales them.

Text performance

Text and HTMLText re-render to a canvas and re-upload to the GPU on every change. Never update them per frame unconditionally:

import { BitmapText, Text } from "pixi.js";

// Wrong: re-renders canvas + GPU upload every frame
app.ticker.add(() => {
  scoreText.text = `Score: ${score}`;
});

// Correct: use BitmapText for frequently changing content
const scoreText = new BitmapText({
  text: "Score: 0",
  style: { fontFamily: "Arial", fontSize: 24, fill: 0xffffff },
});

app.ticker.add(() => {
  scoreText.text = `Score: ${score}`;
});

BitmapText renders from a pre-generated glyph atlas. Updates only reposition quads; no canvas re-render or GPU upload. Use it for scores, timers, counters, and anything that changes frequently.

If you must use canvas Text, guard updates so they only happen when the value changes:

app.ticker.add(() => {
  const next = `Score: ${score}`;
  if (scoreText.text !== next) {
    scoreText.text = next;
  }
});

Text resolution matches the renderer resolution by default. Lower it independently via text.resolution = 1 to reduce GPU memory on high-DPI displays.

Graphics performance

Graphics objects are fastest when their shape doesn't change (transforms, alpha, and tint are fine). Small Graphics (under ~100 points) are batched like Sprites. Complex Graphics with hundreds of shapes are slow; convert them to textures instead:

import { Graphics, Sprite } from "pixi.js";

const complex = new Graphics();
// ... draw complex shape ...

// Render once to texture, use as Sprite
const texture = app.renderer.generateTexture(complex);
const sprite = new Sprite(texture);

Culling

PixiJS skips rendering objects outside the visible area when cullable is set. Disabled by default because it trades CPU cost (bounds checking) for GPU savings. Culling only runs when the CullerPlugin is registered:

import { extensions, CullerPlugin, Culler, Rectangle } from "pixi.js";

extensions.add(CullerPlugin); // before Application.init

// Enable on objects that may be off-screen
sprite.cullable = true;

// Optional: a pre-computed cull rectangle avoids per-frame bounds calculation.
// Without cullArea, the Culler uses the object's global bounds instead.
sprite.cullArea = new Rectangle(0, 0, 800, 600);

// Skip culling an entire subtree (static UI, always visible)
uiRoot.cullableChildren = false;

// Or cull manually without the plugin:
Culler.shared.cull(app.stage, app.renderer.screen);

cullableChildren on a container stops the culler from recursing into its descendants; a large win for static UI panels with many children. Culler.shared.cull(container, rect) runs the same logic manually for custom render pipelines. Use culling when you're GPU-bound; avoid it when CPU-bound, since the per-object bounds check adds overhead.

Resolution and antialias tradeoffs

import { Application } from "pixi.js";

const app = new Application();

// Mobile-friendly: lower resolution, no antialias
await app.init({
  resolution: 1,
  antialias: false,
  backgroundAlpha: 1, // opaque background is faster
});

resolution: 2 quadruples the pixel count. On mobile, this can halve frame rate. Profile to find the right balance.

Low-level rendering optimizations

import { RenderTexture } from "pixi.js";

// single-pass antialiased render texture: discard the MSAA buffer at end of pass
const rt = RenderTexture.create({ width: 1024, height: 1024, antialias: true, transient: true });

// upload only the changed byte range of a large buffer (also works on WebGL)
buffer.update(changedBytes, offsetBytes);

transient: true (WebGPU only) tells the GPU the multisample buffer is scratch memory: it is discarded instead of written back, and tile-based GPUs skip allocating it when renderer.device.extensions.transientAttachment is true. Only use it on textures rendered in a single pass and never re-entered with clear: false or wrapped by a filter. For static custom draw sequences on WebGPU, record a render bundle once and replay it each frame; see the pixijs-custom-rendering skill.

Stagger bulk texture destruction

function staggerDestroy(textures: Texture[], perFrame: number = 5): void {
  let index = 0;
  const ticker = app.ticker;

  const destroy = () => {
    const end = Math.min(index + perFrame, textures.length);

    for (let i = index; i < end; i++) {
      textures[i].destroy(true);
    }
    index = end;

    if (index >= textures.length) {
      ticker.remove(destroy);
    }
  };

  ticker.add(destroy);
}

Destroying many textures in one frame causes a freeze. Spread the cost across frames.

Filters and masks cost

• Set container.filterArea = new Rectangle(x, y, w, h) when you know the bounds. Without it, PixiJS measures bounds every frame.
• Release filter memory: container.filters = null.
• Full-screen filter textures are pooled at screen size, but padding (the BlurFilter default) pushes the request past the screen and up to the next power of two. Set repeatEdgePixels: true on full-screen blurs to keep them screen-sized.
• Mask cost (cheapest to most expensive): axis-aligned Rectangle masks (scissor rect) < Graphics masks (stencil buffer) < Sprite/alpha masks (filter pipeline). Hundreds of masks will slow things down regardless of type; prefer rectangle masks when bounds are axis-aligned.
• Set interactiveChildren = false on containers with no interactive children.
• Set hitArea on large containers to skip recursive child hit testing.

Safe destroy order

Remove from scene before destroying:

parent.removeChild(sprite);
sprite.destroy();

Destroying while the render pipeline still holds a reference causes null-pointer crashes. If destruction must happen mid-frame, defer it:

app.ticker.addOnce(() => {
  parent.removeChild(sprite);
  sprite.destroy();
});

Common Mistakes

[CRITICAL] App destroy without releaseGlobalResources

Wrong:

app.destroy();
const newApp = new Application();

Correct:

app.destroy({ releaseGlobalResources: true });
const newApp = new Application();

Without this flag, stale pooled batches and textures from the old app persist in global pools and get reused by the new app, causing flickering and visual corruption.

[HIGH] Interleaving object types in scene graph

sprite / graphic / sprite / graphic = 4 draw calls. sprite / sprite / graphic / graphic = 2 draw calls.

Group same object types together in the child order to minimize batch breaks. Same applies to blend mode ordering.

[HIGH] Destroying and recreating objects instead of recycling

Destroy/recreate is expensive: it deallocates GPU resources, triggers garbage collection, and requires fresh GPU uploads. Reuse objects by updating texture, position, visible, and other properties. Use an object pool pattern for frequently spawned/despawned entities.

[HIGH] Loading many individual textures instead of spritesheets

Each separate texture consumes its own GPU memory slot and breaks batching when the per-batch texture limit is reached. Spritesheets consolidate textures into atlases. Also avoid textures exceeding 4096px on either axis, as they fail on some mobile GPUs.

[HIGH] Updating Text or HTMLText every frame

Each update re-renders the full string to a canvas and uploads to the GPU. At 60fps this creates massive overhead. Use BitmapText for dynamic content (scores, timers, counters). If canvas Text is required, only update when the value actually changes. Source: src/docs/concepts/performance-tips.md

[HIGH] Using complex Graphics instead of textures

Hundreds of complex Graphics objects are slow to render. Small Graphics (under ~100 points) batch efficiently like Sprites, but complex ones do not. Render complex static shapes to a texture with renderer.generateTexture() and display as a Sprite. Source: src/docs/concepts/performance-tips.md

[MEDIUM] Not staggering bulk texture destruction

Destroying dozens of textures in a single frame causes a visible freeze. Spread destruction across multiple frames (e.g., 5 per frame via a ticker callback). Source: src/docs/concepts/garbage-collection.md

[MEDIUM] Not using PrepareSystem for large scenes

Without renderer.prepare.upload(), textures upload to the GPU on first render, causing frame hitches. For loading screens or scene transitions, upload before displaying. Requires import 'pixi.js/prepare' (not included even in the default bundle; always import it explicitly). Source: src/prepare/PrepareSystem.ts

[MEDIUM] Using high resolution or antialias without profiling

resolution: 2 quadruples the pixel count. antialias: true adds GPU cost. Both degrade performance on mobile devices. Always profile on target hardware before enabling. Source: performance-tips.md

API Reference

• GCSystem
• TextureGCSystem
• RenderableGCSystem
• PrepareSystem
• Culler
• CullerPlugin
• Pool