Skip to main content

Bitmaps, images, and large objects

Compressed file size is not display memory. A 4000 × 3000 ARGB_8888 bitmap is roughly:

4000 × 3000 × 4 bytes ≈ 48 MB

Decoding several such images can exhaust memory even when each JPEG is only 2 MB.

Decode for the display target​

Image loading libraries provide pooling, sampling, lifecycle cancellation, disk/memory caches, EXIF handling, and transformations. Supply requested dimensions/content scale rather than decoding the original into a 96 dp thumbnail.

AsyncImage(
    model = ImageRequest.Builder(LocalContext.current)
        .data(photo.url)
        .size(256, 256)
        .crossfade(true)
        .build(),
    contentDescription = photo.description,
)

Exact APIs depend on library/version; the principle is target-sized decoding.

Pixel formats and fidelity​

ARGB_8888 is commonly 4 bytes/pixel. Lower precision formats may reduce memory but can degrade gradients/transparency/color and are not universally appropriate. Hardware bitmaps can reduce Java heap/copy cost but have mutability/software-rendering/readback limitations. Measure the real pipeline.

Lists and cancellation​

  • Stable keys prevent wrong-image reuse/state churn.
  • Cancel off-screen requests with lifecycle/list ownership.
  • Avoid unique transformed bitmap variants per scroll position.
  • Preload only a bounded near-future window.
  • Do not retain Activity through image callbacks/targets.
  • Placeholder/error sizes should avoid layout jumps.

Large binary data​

Never put bitmaps/byte arrays in Bundle, SavedStateHandle, Intent, WorkManager Data, Room row blobs by default, or a global StateFlow. Persist/stream them and pass stable URI/ID. Binder transaction limits and memory copies can crash or stall the app.

Processing pipeline​

content URI/file -> metadata bounds -> sampled decode/stream
                 -> transform one bounded region
                 -> encode to partial file
                 -> validate -> atomically publish

For upload, stream encoded file rather than decode to bitmap unless transformation is needed. For region/zoom viewers, use tiled/region decoding. Guard pixel dimensions against decompression bombs before allocation.

Mistakes​

  • ❌ BitmapFactory.decodeStream() original for a tiny avatar.
  • ✅ inspect bounds and request/downsample to display size.
  • ❌ convert image to Base64 String and keep it in state.
  • ✅ store file/URI and stream it.
  • ❌ fix OOM only by adding largeHeap=true.
  • ✅ measure retention/decoded sizes/cache and reduce peak working set.
  • ❌ recycle a bitmap still used by UI/library.
  • ✅ let correct owner/library manage lifecycle.

Sources​

Interview question​

Why can a 500 KB photo cause OOM? Encoded size does not determine decoded memory; pixel dimensions × bytes per pixel plus copies/caches/transforms determine peak working set.