Answer Document Question (Pix2Struct)

@google1 VersionGenerate

Visual document question answering with Pix2Struct: takes a document Image and a natural-language String question, renders the question onto the page and decodes a free-text String answer in one pass, without a separate OCR stage.

How it fits

Typical backends

Answer a fixed question about a page fetched over HTTP and publish the answer as JSON for a downstream service.

Render a PDF page, ask a question about it, and return the answer over HTTP.

Caveats

I/O contractThe question is rendered as a header onto the page image before patch extraction, so the model answers directly from pixels with no OCR step; legible page resolution matters more than any text layer.
FallbackA blank page or an empty question short-circuits to an empty String answer, and an inference failure on a single message is logged and also returns an empty answer rather than crashing the stream.
AccuracyAnswers are short extractive spans; the lighter default checkpoint trades accuracy for speed, while the larger DocVQA checkpoint improves accuracy at higher latency and memory.
LatencyThe answer length cap bounds decoding per call, and the model loads once at startup so the first call on a fresh container pays the download and load cost.
CompatibilityRuns on CPU but is materially faster on a GPU, where half precision is used automatically; a GPU host is recommended for interactive latency.

Versions

19c12bfclatestdefault
Automated release
6/29/2026