docs/reference/native-image.mdx
docs/reference/native-image.mdxBrowse 89 files
17,460 bytes
Token encoding: o200k_base
Snapshot ac753b4
title: NativeImage description: Decode, inspect, transform, share, and dispose native PNG, JPEG, WebP, GIF, and RGBA pixel data skill: entry: true intents: [native-image, image-decode, png, jpeg, webp, gif, rgba, pixels, resize]
NativeImage
The Image component owns display and terminal-protocol selection. NativeImage owns image
decode, transformed pixels, and native image handles.
This advanced API is for code that must inspect or transform pixels before display. Use the Image component to load and display an encoded source without direct pixel ownership.
Every materialized image uses top-left, straight-alpha, sRGB RGBA8 pixels. The API is supported in the Bun and Node.js runtimes listed in Runtime and platform support.
Load and dispose an image
import { NativeImage } from "@opentui/core"
const image = await NativeImage.load("./image.webp")
try {
const raw = image.raw("rgba8")
console.log(raw.width, raw.height, raw.stride, raw.data)
} finally {
image.dispose()
}
raw() returns an owned JavaScript copy. Disposing image does not invalidate that copy.
Accepted sources
| API | Accepted input |
|---|---|
NativeImage.load(source, options?) | Path string, file:, HTTP(S), blob:, or data: URL, URL, Blob, Response, Uint8Array, or ArrayBuffer |
NativeImage.decode(data) | Encoded Uint8Array or ArrayBuffer |
NativeImage.fromRgba(pixels, width, height, stride?) | Straight-alpha sRGB RGBA8 Uint8Array |
NativeImage.fromPixels(pixels, width, height, options?) | sRGB RGBA8 or BGRA8 Uint8Array, with straight or opaque alpha |
imageInfo(data) | Encoded Uint8Array or ArrayBuffer |
Format detection reads encoded bytes. It does not trust a filename, URL suffix, response header, or file extension.
decode(), fromRgba(), and fromPixels() copy your data into native memory. imageInfo() returns metadata
without retaining an image handle. It can decode pixels when validation needs pixel inspection.
For fromRgba() and fromPixels(), stride is the byte distance between row starts. It defaults to width * 4 and
must fit a complete row. The input view must contain at least (height - 1) * stride + width * 4 bytes. Padding after
the final row is optional.
Pixel import options
interface PixelImportOptions {
stride?: number
format?: "rgba8" | "bgra8"
alpha?: "straight" | "opaque"
colorSpace?: "srgb"
}
fromPixels() defaults to format: "rgba8", alpha: "straight", and colorSpace: "srgb". It copies rows from a
top-left origin into packed native RGBA8 storage. The same pass converts BGRA to RGBA and detects transparency.
alpha: "opaque" ignores source alpha and writes 255 to every destination alpha channel. It does not unpremultiply
RGB values. The API does not support premultiplied alpha or color spaces other than sRGB.
Pass a subarray directly to import part of a mapped buffer. The native call does not retain or change the view. Keep the source stable during the call. After the call returns, you can reuse, detach, or unmap the buffer.
const image = NativeImage.fromPixels(mappedPixels, width, height, {
format: "bgra8",
stride: bytesPerRow,
alpha: "opaque",
})
The image reports format: "raw-rgba" and colorStatus: "explicit-srgb", even for BGRA input.
Load options
interface ImageLoadOptions {
signal?: AbortSignal
fetch?: (input: URL, init?: RequestInit) => Promise<Response>
}
signal cancels file, response-body, or fetch acquisition. An abort throws signal.reason. fetch replaces
globalThis.fetch for fetched URLs.
Paths, blobs, and response bodies are fully buffered before native decode. load() rejects a non-success HTTP status
before it reads the body.
Format behavior
| Format | Decode behavior |
|---|---|
| PNG | Supports alpha, sRGB chunks, supported cICP, RGB or grayscale ICC v2/v4 profiles, and EXIF orientation |
| JPEG | Produces opaque RGBA8 and applies EXIF orientation |
| WebP | Supports lossy, lossless, and alpha images. Animated WebP is rejected. |
| GIF | Decodes the first displayed frame on the logical canvas |
| Raw pixels | Imports caller-supplied sRGB RGBA8 or BGRA8 into straight-alpha RGBA8 pixels |
NativeImage does not expose animation frames or timing. GIF becomes one image. Animated WebP reports an unsupported
feature.
Metadata
image.info() and imageInfo() return ImageInfo:
| Field | Meaning |
|---|---|
width, height | Decoded and orientation-corrected dimensions |
sourceWidth, sourceHeight | Original input dimensions retained through derived images |
format | "png", "jpeg", "webp", "gif", or "raw-rgba" |
colorStatus | "explicit-srgb" or "assumed-srgb" |
orientation | Encoded orientation from imageInfo(). A decoded image reports 1. |
hasAlpha | Whether decoded pixels contain transparency |
explicit-srgb means the source supplied supported color metadata or the image came from explicit RGBA pixels.
assumed-srgb means OpenTUI treated untagged source values as sRGB.
For PNG, supported sRGB cICP takes precedence. Otherwise, iCCP takes precedence over sRGB, gAMA, and cHRM.
Unsupported cICP falls through to lower-priority metadata. OpenTUI converts supported RGB or grayscale ICC monitor
profiles to sRGB with Little CMS.
Pixel access
| API | Result and ownership |
|---|---|
raw(format = "rgba8") | Allocates and returns a copied RawImage in "rgba8" or "bgra8" order |
copyTo(destination, options?) | Copies into an existing Uint8Array with optional stride and format |
takeRaw() | Transfers exclusive ownership of native RGBA8 pixels to an OwnedRawImage |
width, height | Read decoded dimensions while the handle is valid |
ptr | Read the opaque native ImageHandle while the handle is valid |
Both raw result types include data, width, height, stride, format, colorSpace: "srgb", and
alpha: "straight".
copyTo() defaults to RGBA8 with stride width * 4. A custom stride must fit a row. The destination must fit all rows.
Transfer native pixels
const image = await NativeImage.load(new Blob([encodedImage]))
try {
const raw = image.takeRaw()
try {
consumeRgba(raw.data, raw.width, raw.height, raw.stride)
} finally {
raw.dispose()
}
} finally {
image.dispose()
}
takeRaw() consumes the NativeImage. Later image access throws, while image.dispose() remains a safe no-op.
OwnedRawImage.data aliases native memory. Keep the owner alive while any consumer uses the view.
OwnedRawImage.dispose() is idempotent and required. It frees the native allocation and invalidates data.
takeRaw() requires one exclusive native reference. It throws while another retained handle or a native buffer keeps
the image alive. Dispose extra handles and clear or destroy those buffers first.
Share or copy a handle
| Method | Behavior |
|---|---|
retain() | Returns an independently disposable handle to the same native image without copying pixels |
clone() | Returns a new native image with copied image storage |
dispose() | Releases one handle. The call is idempotent. |
Dispose every retained, cloned, decoded, and transformed handle separately. Disposing one retained handle does not invalidate another retained handle.
Reuse pixel storage
NativeImagePool reuses a fixed number of native pixel buffers, called slots, for frames of the same size.
Each successful publication copies the pixels into a free slot and returns an immutable NativeImage.
import { ImageRenderable, NativeImagePool } from "@opentui/core"
const pool = new NativeImagePool({ width: 640, height: 480, capacity: 3 })
const view = new ImageRenderable(renderer, { width: 80, height: 24 })
renderer.root.add(view)
function onPixels(pixels: Uint8Array): boolean {
const frame = pool.publishRgba(pixels)
if (!frame) return false
try {
view.source = frame
} finally {
frame.dispose()
}
return true
}
// Stop the producer before teardown.
view.destroy()
pool.dispose()
The ImageRenderable.source setter retains the frame synchronously. It requests a render when the load settles, so
you do not need to call requestRender().
Each publication creates a new JavaScript wrapper and native handle. This prevents source equality checks and terminal-protocol caches from hiding a new frame.
new NativeImagePool({ width, height, capacity? })fixes the dimensions and slot count. Capacity defaults to3and must be an integer from1through8. Native image dimension and memory limits apply on the first publication.publishRgba(pixels, stride = width * 4)copies top-left, straight-alpha, sRGB RGBA8 bytes without retaining caller memory. It accepts row padding. The input must include every row's pixels, but padding after the final row is optional.publishPixels(pixels, options?)accepts the samePixelImportOptionsasNativeImage.fromPixels(). It converts RGBA8 or BGRA8 with straight or opaque alpha directly into a reusable slot, without a JavaScript staging buffer.- The pool allocates slots only as needed, up to
capacity * width * height * 4bytes of pixel storage. Reuse does not allocate another pixel buffer. Encoding and rendering can allocate separately. - The pool can reuse a slot only after all published handles and native buffer placements release it. Retained images, renderable sources, and buffer snapshots keep their original pixels. Dispose frames when you no longer need them.
- Both publication methods return
nullif every slot is still in use. They do not allocate pixel storage, queue, block, or replace a frame. Drop the incoming frame or retry after consumers release a slot. Invalid input throws instead of returningnull. - Both methods finish synchronously, with no acquisition token or pending operation to cancel. To cancel input, stop calling them.
dispose()releases the pool's references and rejects future publications. Repeated calls are safe. Published frames remain valid and still need disposal.- Pools cannot resize. Create a new pool for new dimensions. Then dispose the old pool.
takeRaw()cannot transfer a frame while the pool owns its slot. Useraw()orcopyTo(), or dispose the pool first.
Transform images
Operations do not mutate the source. Each successful operation returns a new NativeImage.
| Method | Behavior |
|---|---|
resize({ width?, height?, kernel? }) | Resizes to positive dimensions. One omitted dimension preserves aspect ratio. |
extract({ left, top, width, height }) | Crops an in-bounds pixel rectangle |
extend(options = {}) | Adds top, right, bottom, and left padding |
rotate(90 | 180 | 270) | Rotates clockwise |
flip() | Flips vertically |
flop() | Flips horizontally |
composite(overlay, options = {}) | Composites an overlay in linear light |
Resize kernels are "area", "default", "triangle", "cubic-bspline", "catmull-rom", "mitchell", and
"nearest". The default is "area".
extend() defaults all sides to zero and the background to transparent [0, 0, 0, 0]. Each background channel is an
integer from 0 through 255.
Composite options default to left: 0, top: 0, blend: "source-over", and opacity: 1. Blend modes are
"source-over", "source", and "destination-over". Opacity must be finite and in 0..1. Negative offsets clip
the overlay to the base image.
const source = await NativeImage.load("photo.jpg")
try {
const thumbnail = source.resize({ width: 320 })
try {
useImage(thumbnail)
} finally {
thumbnail.dispose()
}
} finally {
source.dispose()
}
Encoded PNG retention
ensureEncodedPng() makes encoded PNG data available for low-level native consumers. Raw and transformed images can
need an encode at this point. The method returns void and leaves the image usable.
Opaque, orientation-free PNG input can retain its original encoded bytes without decoding pixels immediately. A direct unchanged Kitty placement can use those bytes. Any pixel read, pixel operation, Sixel or block rendering, or Kitty crop, resize, or opacity operation materializes pixels.
This lazy path means corrupt PNG pixel data can pass initial metadata validation and fail when the first pixel path runs.
Errors and limits
ImageLoadError covers source acquisition. It exposes code, source, and optional HTTP status.
| Code | Source failure |
|---|---|
file-read | Path or file URL read |
network | Fetch or response-body read |
http-status | Non-success HTTP status |
unsupported-url-scheme | URL outside file, HTTP(S), blob, or data support |
ImageError covers native decode and operations. It exposes numeric status and one of these codes:
invalid-handle,unsupported-format,unsupported-color-space, ormalformed-datadimension-limit,memory-limit,invalid-argument, orout-of-memoryoutput-too-small,internal-error, orunsupported-featurebusyfor retained native pool slots. Pool publication methods returnnullinstead of throwing this status.
JavaScript option and geometry validation uses TypeError and RangeError. All image methods except dispose() throw
after disposal or transfer.
| Limit | Value |
|---|---|
| Encoded input | 64 MiB |
| Width or height | 16,384 pixels |
| Total pixels | 25 million |
| RGBA storage per image | 100 MiB |
| Decompressed ICC profile | 8,000,000 bytes |
OpenTUI keeps the eight most recently used validated ICC profiles and sRGB transforms. The cache key includes complete decompressed profile bytes and RGB or grayscale mode. The native library clears the cache when its final client releases it.
Next
- Image owns normal display and protocol choice.
- FrameBuffer displays an existing or transformed image.
- Buffer API defines
drawImage()placement ownership. - Lifecycle and cleanup defines native-handle cleanup.
- Runtime and platform support lists native targets.
- API and symbol index lists every image export.
Referenced from SKILL.md
Source excerpt starting at line 103.SKILL.mdView in source ↗103| `env`, `environment`, `configuration`, `flags` | `docs/reference/env-vars.mdx` |104| `native-image`, `image-decode`, `png`, `jpeg`, `webp`, `gif`, `rgba`, `pixels`, `resize` | `docs/reference/native-image.mdx` |
Source excerpt starting at line 141.141- `docs/reference/env-vars.mdx`142- `docs/reference/native-image.mdx`