ZoomCap: Evidence-Routed Visual Zooming for Recoverable Visual Information Transfer

1 Introduction

• Prior caption measures emphasize fluency or length, but the paper frames the central question as whether transferred visual evidence remains verifiable, recoverable, and useful.A single global pass can miss small or peripheral evidence, while fixed high-resolution crops may spend budget on empty background.Source: §1 · p. 2
• From a draft, ZoomCap conditionally routes knowledge search, OCR, omitted-region recovery, and deeper zooming, then fuses supported, non-duplicate evidence and derives relations from geometry.Deeper coverage is limited to the highest-scoring eligible region.Source: §1 · p. 2
• ZoomCap transfers 52.4 facts per image (1.7× the strongest baseline), earns an 81.9% human tie-adjusted preference score, and improves the controlled 11-benchmark mean by 1.31–2.38 points.The preference comparison is against Qwen3.5-397B-A17B captions; the controlled mean comparison is across three backbones.Source: §1 · p. 2
• FactCycle scores caption-derived atomic claims against the source image for truthfulness and a fixed reconstruction for realizability, separating fabrication from generator loss claim by claim.The measured facts include entities, attributes, relations, counts, small objects, and actions.Source: §1 · p. 2
• ZoomCap-450K connects caption quality to reusable supervision through controlled continual pre-training and production adoption in general and medical domains.Production results include gains on document/OCR, perception, and medical VQA evaluations.Source: §1 · p. 2
• Figure 1 shows five targets retained by ZoomCap but partial or missing in ScaleCap and Gemini-3.1-Pro, alongside 1.45× total characters and 1.58× P99 length versus ScaleCap.These comparisons illustrate evidence preservation as well as greater caption length.Source: Figure 1 · p. 3

Caption supervision and caption-data generation

• Synthetic-caption pipelines use strong multimodal models, vision experts, web knowledge, or scalable workflows; ZoomCap conditionally routes search, OCR, geometry checks, and crops from a grounded draft.Its budget is spent where the draft remains unexplained.Source: §2 · p. 3

Fine-grained evidence for faithful detailed captioning

• Region grounding, detailed-caption datasets, retrieval, visual search, and scene graphs address local evidence; ZoomCap unifies these channels under a finite budget and admits spatial facts only with geometric support.This keeps spatial phrases auditable at full coverage.Source: §2 · p. 3

Caption evaluation as factual and cycle-consistent measurement

• Prior caption metrics assess alignment, factuality, downstream QA, or reconstructed-image similarity; FactCycle instead scores atomic claims on blind source-image and fixed-reconstruction anchors.The two anchors distinguish truthfulness, realizability, and generator loss claim by claim.Source: §2 · p. 3

3 ZoomCap Method

• ZoomCap frames captioning as budgeted evidence acquisition: a global draft identifies what is covered, routes needed observations, and conditions final fusion.The objective balances coverage, factuality, spatial faithfulness, recoverability, and downstream usefulness under a finite observation budget.Source: §3 · p. 4, §3.1 · p. 4

3.2 Draft-Conditioned Evidence Routing

• Three draft-conditioned streams repair different gaps: world-knowledge search, OCR for text-structured images, and zoom-in for under-described detail-rich regions.Search may name or explain entities but cannot override the visual scaffold; OCR preserves reading order by summarizing recognized text.Source: §3.2 · p. 4

3.3 Coverage-Aware Visual Zooming

• Coverage-aware zooming combines quadrant, semantic, and draft-relative gap regions, removes overlap, and expands only the highest-scoring parent approved by complexity rules and an agent.Gap scoring measures what lies outside regions already explained by the draft; the complexity score considers entity count, small-object count, clipped-entity density, and area ratio.Source: §3.3 · p. 5

3.4 Geometry-Grounded A-SPOE Relations

• A-SPOE represents grounded entities as attribute-bearing graph nodes and admits spatial relations only when localized box geometry supports them.The geometric test uses normalized horizontal and vertical center gaps and a margin, keeping spatial phrases grounded before they enter free-form text.Source: §3.4 · pp. 5–6

3.5 Draft-Anchored Gated Fusion

• Draft-anchored fusion treats the draft as the backbone and admits candidate statements only when they are novel, visually consistent, and non-redundant.Admitted statements retain provenance, while uncertain local observations are not promoted into confident global claims.Source: §3.5 · p. 6

3.6 FactCycle: Two-Anchor, Fact-Level Measurement

• FactCycle scores the same blind-extracted claims against source truthfulness and reconstruction realizability, separating transferred claims from fabricated or unrendered claims.Its four cells include transferred claims (true and rendered), specious claims (fabricated yet rendered), and unrendered fabrication; true but unrendered claims are treated as generator loss.Source: §3.6 · pp. 6–7
• Unlike reconstruction-only cycle metrics, FactCycle does not credit fabricated-but-rendered claims as caption evidence and uses source–reconstruction similarity as a judge-independent global check.Fact-level verdicts attribute unsupported content to entities, attributes, relations, counts, small objects, or actions.Source: §3.6 · p. 7

4 Experiments

• The evaluation tests whether captions preserve recoverable source-grounded facts, identifies which routed components contribute, and measures downstream training transfer; captions generally use a Qwen3.5-35B-A3B teacher.Source: §4 · p. 7

4.1.1 FactCycle: Fact-Level Truth and Realizability

• 52.4 transferred facts per image: ZoomCap reaches 1.7× ScaleCap and 3.5× ShareGPT4V, with higher truthfulness (93.8% vs. 88.9% and 81.4%) and reconstruction similarity (0.696 vs. 0.620 and 0.566).The fixed-generator evaluation uses 250 paired images; ZoomCap leads every rate column despite captions 1.45× longer than ScaleCap.Source: §4.1.1 · pp. 7–8, Figure 5 · p. 7, Table 1 · p. 8

4.1.2 Source-Grounded Preference

• 81.9%: humans prefer ZoomCap over Qwen3.5-397B-A17B on PrefBench-456, and all three VLM judges agree; knowledge richness is strongest and coverage improves consistently.Figure 6 identifies knowledge richness and information coverage as the largest gains.Source: §4.1.2 · pp. 8–9, Figure 6 · p. 8

4.2 What Produces the Gain?

• 71.2% to 28.8%: complexity-gated zooming beats the unrestricted pyramid head-to-head, using 20.2% fewer local crops and 39.8% fewer second-level crops.In the evidence-stream ablation, retrieval raises knowledge richness from 46.7 to 92.3 and the score to 74.2.Source: Table 2 · p. 8, §4.2 · p. 9

4.3 Transfer to VLM Pre-Training

• 1.31–2.38 points: ZoomCap-450K beats the strongest competing 450K corpus on the 11-benchmark mean across three backbones; production gains include 1.4 document/OCR points, 1.5 perception points, and a 3.04-point medical five-VQA mean gain over the reference.The medical same-image Gemini-3.1-Pro comparison is a 1.28-point gain.Source: Table 5 · p. 9, Table 3 · p. 9, Table 4 · p. 9

• ZoomCap conditions on a global draft to acquire knowledge, OCR, and zoom-in evidence, then adds geometry-grounded relations and fuses only novel, consistent, non-duplicate evidence.Source: §5 · p. 10
• FactCycle scores caption-derived atomic claims against source-image truthfulness and fixed-reconstruction realizability, separating fabrication from generator loss claim by claim.Source: §5 · p. 10
• Medical-domain results are benchmark evaluations of pre-training recipes, not clinical validation; diagnosis or treatment use requires domain-specific validation and oversight.Source: Ethics statement · p. 10
• The controlled training comparisons use one fixed seed, while internal production datasets cannot be redistributed, limiting full public reproducibility.Source: Reproducibility statement · p. 10

• The workflow records drafts, search context, OCR summaries, grounded boxes, patch coordinates and captions, A-SPOE facts, routing decisions, latency, and token use for auditing and ablation.The resulting image-caption pairs are used directly as supervision in the controlled CPT and production-adoption studies.Source: A Evaluation and Training Protocol… · p. 15
• The evaluation has caption-, mechanism-, and data-level tests: FactCycle and human/multi-judge comparisons, evidence-stream and zoom-budget ablations, and controlled or production pre-training studies.Source: A Evaluation and Training Protocol… · p. 15
• FactCycle-250 samples 50 images from each of five scene categories after deduplication; screenshots, documents, charts, collages, artwork, and illustrations are excluded, so this experiment does not test OCR.All captions are reconstructed with the same Qwen-Image-2512-Lightning generator and seed; prompt truncation affects ZoomCap more often, but the reported ordering and significance tests persist on the untruncated subset.Source: A Evaluation and Training Protocol… · p. 15
• FactCycle uses a text-only extractor for atomic claims and independent judges who see either the source image or reconstruction with anonymized claims; per-image judge scores are averaged before aggregation.Source: A Evaluation and Training Protocol… · p. 15
• PrefBench-456 uses randomized two-alternative forced choice, six annotators, and VLM judges; tie-adjusted scores count wins plus half the ties, so paired sides sum to 100.Source: A Evaluation and Training Protocol… · p. 15
• Controlled CPT changes the 450K caption corpus while holding initialization, other training data, resolution, and evaluation protocol fixed within each backbone block; all runs use seed 42.The recipe includes projector initialization, one CPT epoch, and one instruction-tuning epoch.Source: A Evaluation and Training Protocol… · p. 15

• Coverage gaps are rectangles bounded by image and box edges that do not intersect the union of boxes already explained by the draft; near-full-image boxes and extremely narrow gaps are rejected.A second gap may be selected after adding the first to the covered set.Source: B Routing and Zoom Implementation… · p. 16
• Second-level zoom eligibility uses entity counts, small-object counts, clipped entity density, or entity coverage; an agent can veto uninformative nominations, and only the highest-scoring approved parent proceeds to fusion.Source: B Routing and Zoom Implementation… · p. 16
• Relations are assigned from normalized horizontal and vertical box-center gaps after filtering overlapping or near-coincident pairs; candidates are ranked geometrically and capped at eight relations.Source: B Routing and Zoom Implementation… · p. 16
• Fusion receives the draft and ordered patch descriptions and appends only a new visual-details section, excluding already stated content and encyclopedic knowledge; geometry-derived relations retain their supplied direction.Source: B Routing and Zoom Implementation… · p. 16

C Multi-Judge Cross-Validation

• Figure 8 contrasts localized attribute grounding and small-text recovery, marking visually incorrect spans red and image-supported spans green.Its representative cases show ZoomCap preserving local evidence that Qwen3.5-397B-A17B reverses or misreads.Source: Figure 8 · p. 16
• Table 6 reports human tie-adjusted preference scores, with 81.9 for the General verdict; its Overall verdict aggregates six annotators’ dimension rates.The table also includes each judge’s Overall verdict; the supplied passage does not give those values.Source: Table 6 · p. 17

D More Qualitative FactCycle Reconstructions

• Figure 9 adds fourteen paired FactCycle reconstruction sources, using the same fixed generator and caption sources as Figure 7.The figure shows the fourteen pairs omitted from Figure 7.Source: Figure 9 · p. 17

E Additional FactCycle Breakdowns

• FactCycle scores each caption’s blind-extracted claim set against the source image for Truthfulness τ and a fixed-generator reconstruction for Realizability ρ.The aggregation paragraph defines the denominator and handling of split reconstruction judgments for quadrant rates and FCtrans.Source: E Additional FactCycle Breakdowns · pp. 17–18
• Changing aggregation order shifts quadrant shares by at most 3.0 points and FCtrans by 4.1, while preserving every ordering and both significance tests.The alternative averages judge scores per claim before intersecting anchors and treats a claim as decidable if either judge decided it.Source: E Additional FactCycle Breakdowns · p. 18
• Every scene bucket preserves the overall FCtrans ordering; Table 7 reports two-anchor scoring across five buckets of 50 images each.Its transferred, specious, and unrendered-fabrication shares cover claims decidable on both anchors, with the true-but-unrendered generator cell excluded from caption metrics.Source: E Additional FactCycle Breakdowns · p. 18, Table 7 · p. 18
• ZoomCap has the lowest specious share in all seven fact categories and the highest transferred share in six; ScaleCap leads transferred share for small objects, 85.4 vs. 84.2%.Table 8 reports shares for claims decidable on both anchors within each category.Source: E Additional FactCycle Breakdowns · p. 18, Table 8 · p. 18
• For ShareGPT4V-450K, 24.1% of decidable action claims and 20.1% of count claims are specious, meaning fabricated yet rendered.The true-but-unrendered generator share is slightly elevated for ZoomCap on counts (8.5%) and spatial relations (5.6%); the authors attribute this to generator loss, not caption error.Source: E Additional FactCycle Breakdowns · pp. 18–19, Table 8 · p. 18
• ZoomCap’s realizability is 92.3% on decidable claims and 88.8% when unverifiable verdicts count against it; realizability alone cannot rank caption quality.The 3.5-point gap equals ZoomCap’s unverifiable share, and decidable-only realizability is highest for the shortest captions in every scene bucket.Source: E Additional FactCycle Breakdowns · p. 19

F Corpus Statistics

• Table 10 describes caption length, entity density, and image sources for the equal-scale 450K datasets; downstream utility is evaluated in Table 5.The supplied passages give no specific Table 10 values.Source: F Corpus Statistics · p. 19, Table 10 · p. 19

G General-Domain Adoption Details

• +1.4 on the nine-benchmark Doc & OCR suite mean and +1.5 on the perception suite mean were reported for Ling-Tiny-VL-v3.The knowledge-VQA suite gain was +0.7; the document/OCR gains include OmniDocBench +9.8 and OCRBenchV2-b +2.2, while perception gains include MMVet +11.8 and HallusionBench +1.4.Source: G General-Domain Adoption Details · p. 19
• Figure 10 shows the Doc & OCR, OmniDocBench, and perception suite means across shared 40–55B-token checkpoints, with gains stable across checkpoints.The caption contrasts the baseline mixture with the mixture containing ZoomCap captions.Source: Figure 10 · p. 20

H Medical-Domain Adoption Details

• +3.04 on average across five medical benchmarks for ZoomCap captions, versus +1.76 for same-image Gemini-3.1-Pro captions.Both augmented runs used 64 GPUs for approximately 50B tokens, versus 128 GPUs and approximately 100B tokens for the reference; both outperformed it.Source: H Medical-Domain Adoption Details · p. 20
• ZoomCap captions changed the internal hallucination monitor by −0.5, versus −1.5 for Gemini-3.1-Pro; the diagnostic does not identify the cause of the difference.The study adds 2.64M medical-image captions as a dedicated group covering 20% of the mixture.Source: H Medical-Domain Adoption Details · p. 20

Efficiency

• Observation count, rather than any single local call, dominates cost, so budgeted zooming reduces unnecessary second-level patches and local-hallucination risk.The normalized per-image cost accounts for calls by workflow stage and token usage and prices.Source: I Efficiency and Auditability · p. 20
• $0.0077 per image for ZoomCap is 3.6× cheaper than ScaleCap and each transferred claim costs ∼5.9× less, while transferring 1.66× more claims.The comparison uses the same service and teacher; halving ScaleCap’s measured input tokens would still leave it ∼2.0× more expensive per image.Source: I Efficiency and Auditability · p. 20
• Concurrent knowledge and grounding stages and five parallel patch workers bring per-image wall-clock time to ∼29 s versus roughly 55 s sequentially.The workflow processes separate images simultaneously, so per-image time follows the critical path rather than the sum of calls.Source: I Efficiency and Auditability · p. 21

Auditability

• Intermediate evidence artifacts make failures traceable to routing, search, OCR, patches, relation facts, or fusion rather than one opaque generation call.Table 11’s cost accounting compares ZoomCap production logs, the official ScaleCap recipe on the same service and teacher, and a logged single-pass flagship run.Source: I Efficiency and Auditability · p. 21, Table 11 · p. 21