Description

A tile retained while hidden by a nonzero GPU byte target remains marked unused in the plugin cache when shown again. A later cleanup can remove its cache entry while the visibility guard skips GPU disposal. This understates estimatedGpuBytes and prevents that tile from being unloaded on a later hide.

The visibility callback updates the renderer cache through tiles.markTileUsed(tile), but not the plugin cache. lruCache.add(tile, ...) does not mark an existing item used again. Proposed fix: also call the plugin cache's lruCache.markUsed(tile) on every visible transition, keeping the renderer cache marking unchanged.

I have two regression tests and an AI-assisted fix prepared as a draft for discussion. The first regression fails against master and passes with the additional mark.

Reproduction steps

1. Register UnloadTilesPlugin({ bytesTarget: 100 }) and use two tile scenes each estimated at 100 bytes. Set an equal-priority comparator for deterministic eviction order.
2. Show tile A, hide it, and let the scheduled cleanup run. The budget retains A as unused.
3. Show A again, then show and hide B. Let cleanup run.
4. On master, A is removed from the plugin cache even though it is visible; its unload callback skips actual disposal and B remains resident.
5. Set bytesTarget to zero and hide A again. A cannot be unloaded because it is no longer tracked. With the fix, B unloads in step 3, A remains counted, and A unloads in step 5.

Code

import { Group } from 'three';
import { TilesRenderer } from '3d-tiles-renderer/three';
import { UnloadTilesPlugin } from '3d-tiles-renderer/three/plugins';

const tiles = new TilesRenderer();
tiles.lruCache.unloadPriorityCallback = () => 0;
const plugin = new UnloadTilesPlugin({ bytesTarget: 100 });
tiles.registerPlugin(plugin);
tiles.dispatchEvent({ type: 'update-before' });
// Controlled byte estimates isolate residency tracking from mesh size.
tiles.calculateBytesUsed = () => 100;
const a = { engineData: { scene: new Group() } };
const b = { engineData: { scene: new Group() } };

tiles.setTileVisible(a, true);
tiles.setTileVisible(a, false);
await Promise.resolve();
tiles.setTileVisible(a, true);
tiles.setTileVisible(b, true);
tiles.setTileVisible(b, false);
await Promise.resolve();

console.log(plugin.lruCache.has(a)); // expected true; master: false
console.log(plugin.lruCache.has(b)); // expected false; master: true

Live example

No hosted browser sandbox. Runnable regression tests and the proposed fix are available in the fork draft: hoyleb#1

In that checkout: npm ci && npx vitest run test/three/UnloadTilesPlugin.test.js. The first regression was verified failing on unmodified master and passing after the fix.

Screenshots & Tile Set

No dataset is needed: the reproduction uses two synthetic Group scenes and controlled byte estimates to isolate the cache lifecycle. No browser GPU benchmark was performed.

Library Version

0.5.3; master at 075b123

Three.js Version

0.185.1