Does higher layout F1 justify changing the default detector?

By Wei Jie Chee · 28 Jul 2026, 00:00 Z

Download printable cheat-sheet (CC-BY 4.0)
Failure ownerTest layout and table claims with narrower evidence.

Higher aggregate layout F1 is insufficient when label coverage, box fragmentation, reading order, and fallback behavior weaken the full pipeline.

Do this first

Compare the missing labels, fragmented boxes, reading order, and fallback cost before changing the detector.

Keep this boundary in view

The result applies to the tested detectors, labels, public-safe pages, and pipeline contract, not every Surya mode.

Inspect the evidence and what would change the conclusion
Evidence layer
Public-safe Surya and Paddle layout comparison
Observed
28 Jul 2026
Sample
Bounded 15-page and 60-page layout diagnostics

What supports it

Surya OCR 2 beat a Paddle detector on aggregate layout F1 in two public-safe tests, while bounded diagnostics exposed operational regressions.

Change our mind if

Surya improves end-to-end extraction on a representative held-out set without losing required labels or increasing fallback cost.

Claim OCR-LAYOUT-SURYA2-V1 / v1, checked 6 Aug 2026

A higher layout score does not automatically mean a better document pipeline. We learned that while testing Surya OCR 2 against a Paddle layout detector on public-safe historical documents.

Surya reached a higher aggregate Region F1 on two labelled samples. It also found more text regions. But it produced fragmented boxes, weaker label agreement on one sample, and incomplete non-text coverage. When we used its boxes to drive downstream OCR, the extra crops often duplicated or split text that was already covered.

The result was useful, but narrower than "Surya beats Paddle." It showed why OCR layout evaluation needs several metrics and a visual audit before a detector becomes the default.

The short version

  • Surya OCR 2 completed a 60-page layout run with no failures or warnings.
  • On that sample, Surya's Region F1 was 0.669536, versus 0.619309 for Paddle.
  • The gain came mainly from stronger text-region recall.
  • Surya emitted no figure or separator regions in that scored view, while Paddle recovered some figures.
  • A second 15-page dataset again favored Surya on Region F1, but lower precision and over-segmentation persisted.
  • Surya-selected crops could be read by FireRed, but many additional crops were fragments or duplicates rather than new document evidence.
  • We kept Paddle as the default detector and Surya as a diagnostic challenger.

What Surya OCR 2 actually does

Surya 2 is a compact document model that supports OCR, layout analysis, reading order, and table recognition. The upstream release describes a roughly 650M parameter model served through vLLM on NVIDIA GPUs or llama.cpp on CPU and Apple Silicon. The public model card reports about 0.7B parameters.

Those upstream claims establish the intended capabilities. They do not establish how Surya behaves on our documents. Our findings below are local reproductions on an RTX 3090 Ti and public-safe labelled samples.

Sources: Surya repository, Surya OCR 2 model card, and Surya 2 release notes.

The layout question we tested

The practical question was not "Which model has the highest score?" It was:

Can Surya replace Paddle as the detector that finds page regions before OCR?

That role requires more than finding text. A useful detector must preserve the regions, labels, and order needed by the downstream workflow without creating excessive duplicate or fragment crops.

We separated four things that are often mixed together:

Local test / OCR-LAYOUT-SURYA2-V1

Test this claim on your documents.

Copy a read-only experiment brief that asks your coding agent to preserve raw evidence, test a negative case, and return a verdict that can disagree with this page.

Follow OCR research via RSS

Preview the experiment brief
Use Codex to test this bounded OCR claim in my local environment.

Source: https://instavar.com/research/ocr/surya-ocr-2-layout-f1-default-detector
Claim ID: OCR-LAYOUT-SURYA2-V1
Claim version: v1, checked 6 Aug 2026
Claim: Higher aggregate layout F1 is insufficient when label coverage, box fragmentation, reading order, and fallback behavior weaken the full pipeline.
Observed evidence: Surya OCR 2 beat a Paddle detector on aggregate layout F1 in two public-safe tests, while bounded diagnostics exposed operational regressions.
Evidence layer: Public-safe Surya and Paddle layout comparison.
Sample: Bounded 15-page and 60-page layout diagnostics.
Boundary: The result applies to the tested detectors, labels, public-safe pages, and pipeline contract, not every Surya mode.
Potential falsifier: Surya improves end-to-end extraction on a representative held-out set without losing required labels or increasing fallback cost.

Safety and privacy:
- Work read-only by default and keep all documents and outputs local.
- Do not upload documents, reveal credentials, or send document contents to an external service.
- Do not install software, download large models, start paid jobs, or change my environment without asking first.
- If the available tools cannot run the comparison safely, stop with a concrete setup plan instead of pretending the claim was tested.

Test method:
1. Inspect the available OCR workflows and record their exact versions, runtime, hardware, and configuration.
2. Help me choose a small representative sample. Include a known difficult page and a negative control with expected text or structure when available.
3. Before running anything, state the primary failure we are testing and the measurement that would expose it.
4. Preserve raw outputs. Compare text fidelity, structure, blank-page behavior, grounding correctness, runtime, and cleanup effort only where those properties can actually be measured.
5. Treat anchor count as an unverified quantity until sampled anchors are checked against the page.
6. Report the result as support, contradict, mixed, or inconclusive. Explain what observation produced that verdict.
7. Name the smallest follow-up test that could change the verdict.
What did your local test say?

Save a coarse verdict in this browser. Instavar receives no document contents through this control.