AI Operations

Node Browser AI

AI-powered transforms. No API key required — all models run locally via ONNX.

AI models are lazy-loaded on first use. Expect a short delay on the first call as the model downloads and initialises. Subsequent calls are fast.

removeBackground()

.removeBackground(): Pipeline

.removeBackground(): Pipeline

Remove the image background using a U2-Net ONNX model. Outputs a PNG with transparent background (alpha channel).

const png = await img('portrait.jpg')
  .removeBackground()
  .toBuffer()

// In browser
const dataUrl = await img(file)
  .removeBackground()
  .toDataURL()

const png = await img('portrait.jpg')
  .removeBackground()
  .toBuffer()

// In browser
const dataUrl = await img(file)
  .removeBackground()
  .toDataURL()

The output format is always PNG when background removal is used, regardless of the downstream format operation. Chain a format conversion if you need a specific output:

await img('photo.jpg')
  .removeBackground()
  .resize(800)
  .png()
  .toBuffer()

await img('photo.jpg')
  .removeBackground()
  .resize(800)
  .png()
  .toBuffer()

smartCrop()

.smartCrop(subject?: 'face' | 'object'): Pipeline

.smartCrop(subject?: 'face' | 'object'): Pipeline

Content-aware cropping using COCO-SSD object detection. Identifies the primary subject and crops to keep it centred.

Prop	Type	Default	Description
subject	`'face' \| 'object'`	`'object'`	Hint for detection model. `'face'` prioritises face regions.

// Crop to 400x400 keeping the main subject
await img('portrait.jpg')
  .smartCrop('face')
  .resize(400, 400, { fit: 'cover' })
  .webp()
  .toBuffer()

// Crop to 400x400 keeping the main subject
await img('portrait.jpg')
  .smartCrop('face')
  .resize(400, 400, { fit: 'cover' })
  .webp()
  .toBuffer()

Chain smartCrop() before resize(). The crop happens first, then the resize scales the result to your target dimensions.

upscale()

.upscale(factor?: 2 | 4): Pipeline

.upscale(factor?: 2 | 4): Pipeline

AI upscaling using ESRGAN. Increases resolution while recovering detail that conventional bicubic interpolation loses.

Prop	Type	Default	Description
factor	`2 \| 4`	`2`	Upscale multiplier. 2x doubles dimensions, 4x quadruples.

// 2x upscale (default)
await img('low-res.jpg').upscale().toBuffer()

// 4x upscale
await img('thumbnail.jpg').upscale(4).toBuffer()

// 2x upscale (default)
await img('low-res.jpg').upscale().toBuffer()

// 4x upscale
await img('thumbnail.jpg').upscale(4).toBuffer()

4x upscaling is memory-intensive. For images above 1MP, prefer 2x and run twice if needed.

Combining AI ops

// Remove background, upscale, export
await img('product.jpg')
  .removeBackground()
  .upscale(2)
  .png()
  .toBuffer()

// Remove background, upscale, export
await img('product.jpg')
  .removeBackground()
  .upscale(2)
  .png()
  .toBuffer()