Resize Images in the Browser Before Sending to Vision Models to Cut Costs

Sending full-resolution phone photos to AI vision models is expensive because these models process images as fixed-size tiles, meaning cost scales with image area. A 12-megapixel phone photo can generate up to 48 tiles, costing roughly 12 times more than a 1024×1024 version of the same image. For most vision tasks like description or classification, resizing to 768–1024 pixels on the long edge is sufficient without any loss in output quality. Text-heavy images are an exception, requiring 1536–2048 pixels or a cropped region at full resolution to preserve readable detail. Developers can use the browser's createImageBitmap API to resize images off the main thread before upload, reducing both cost and latency.