Extract Formula (PaddleX)

How it fits

Typical backends

Caveats

• Hard constraintEach frame is read as a single formula. Feed a tight crop of one formula region, never a full document page; multiple formulas or surrounding text degrade the transcription.
• Hard constraintuse_gpu binds the model to the first CUDA device at startup and requires a real GPU to be present. There is no automatic switch to CPU when GPU is requested.
• I/O contractEvery frame emits exactly one String. A zero-size or sub-stride Image short-circuits to an empty String; a larger blank or flat frame still decodes to a short spurious formula because the recognition head is generative. The output is raw LaTeX that is never rendered or validated as math, so malformed formulas surface only downstream.
• Hard constraintThe model slot MUST bind a PaddleX inference bundle (config.json + inference.json + inference.yml + inference.pdiparams) whose head matches formula_model_name. The bundle loads entirely offline; nothing is fetched from a remote model server at deploy or runtime.
• Accuracyformula_model_name names the recognition head the bound model bundle holds and is validated against the published PP-FormulaNet set; a name outside it fails at startup. The plus-M default is the highest-accuracy English+Chinese head (2560 tokens); the lighter L/S and UniMERNet/LaTeX_OCR_rec heads trade accuracy for throughput.
• State lifetimeAll settings are captured once at startup; changing any of them takes effect only after a redeploy.