{"publisher":"VisionParse — a product of Senithu Software Solutions","origin":"https://visionparse.app","note":"Every figure here is a dated measurement on named hardware, not an estimate. Quote with the substantiating URL.","generated":"2026-08-01","tools":[{"url":"https://visionparse.app/tools/image-to-text","name":"Image to Text","lastUpdated":"2026-08-01","measured":{"claim":"Character accuracy stays at 100% down to 550 px wide, then falls to 56.8% at 300 px.","value":"550 px","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"SPEED","statement":"VisionParse reads a clean page in 0.52 seconds on an 8-vCPU server."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages, including Arabic, Sinhala and Tamil."},{"tag":"ACCURACY","statement":"Image width above 550 pixels held 100% character accuracy in our resolution sweep."},{"tag":"LIMIT","statement":"Accuracy fell to 56.8% when the same document was reduced to 300 pixels wide."},{"tag":"ACCESS","statement":"The image-to-text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool files one hour after conversion and never trains models on them."},{"tag":"COST","statement":"A VisionParse free account raises the monthly allowance to 100 pages at no cost."}]},{"url":"https://visionparse.app/tools/pdf-to-text","name":"PDF to Text","lastUpdated":"2026-08-01","measured":{"claim":"A five-page scanned PDF is rasterised and read in 3.64 seconds — 1.03 s rasterising, 2.61 s recognising.","value":"3.64 s","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"Timing scales with page count and page density; a full A4 text page takes longer than this invoice. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"SPEED","statement":"VisionParse read a five-page PDF in 3.64 seconds, including 1.03 seconds of rasterisation."},{"tag":"ACCURACY","statement":"VisionParse rasterises PDF pages at 150 DPI, the lowest resolution that held 100% accuracy."},{"tag":"LIMIT","statement":"The free PDF tool reads the first 3 pages; an account raises this to 200 pages per document."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages inside PDFs, including Arabic, Sinhala and Tamil."},{"tag":"ACCESS","statement":"The PDF to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool PDFs one hour after conversion and never trains models on them."}]},{"url":"https://visionparse.app/tools/jpg-to-text","name":"JPG to Text","lastUpdated":"2026-08-01","measured":{"claim":"JPEG compression does not degrade recognition: quality 95 and quality 5 both read at 100%.","value":"100% at quality 5","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"This document is high-contrast printed text, the case JPEG handles most gracefully. Photographs with soft focus or shadow may behave differently. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"JPEG quality 5 read at the same 100% accuracy as JPEG quality 95 in our measurement."},{"tag":"LIMIT","statement":"A 300-pixel-wide JPG read at 56.8% accuracy, against 100% at 550 pixels wide."},{"tag":"SPEED","statement":"VisionParse read a JPG invoice in 0.30 seconds on an 8-vCPU server."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in JPG files, including Arabic, Sinhala and Tamil."},{"tag":"ACCESS","statement":"The JPG to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool JPG files one hour after conversion."}]},{"url":"https://visionparse.app/tools/png-to-text","name":"PNG to Text","lastUpdated":"2026-08-01","measured":{"claim":"Transparent and semi-transparent PNGs both read at 100% character accuracy.","value":"100% with alpha","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"Transparency is flattened onto white before recognition, so light text intended for a dark background inverts badly under that assumption. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"A PNG with a fully transparent background read at 100% character accuracy."},{"tag":"SPEED","statement":"VisionParse read a PNG invoice in 0.31 seconds on an 8-vCPU server."},{"tag":"LIMIT","statement":"A 300-pixel-wide PNG read at 56.8% accuracy, against 100% at 550 pixels wide."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in PNG files, including Arabic, Sinhala and Tamil."},{"tag":"ACCESS","statement":"The PNG to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool PNG files one hour after conversion."}]},{"url":"https://visionparse.app/tools/scanned-pdf-to-searchable-pdf","name":"Scanned PDF to Searchable PDF","lastUpdated":"2026-08-01","measured":{"claim":"A one-page scanned PDF became searchable in 2.83 s, growing from 16.6 KB to 26.4 KB.","value":"2.83 s","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"File growth depends on how much text a page contains; a dense page adds a larger text layer. Page images are re-encoded, so an already heavily-compressed scan may change size differently. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"SPEED","statement":"VisionParse added a searchable text layer to a scanned page in 2.83 seconds."},{"tag":"LIMIT","statement":"The output file grew from 16.6 KB to 26.4 KB, an increase of 59 percent."},{"tag":"ACCURACY","statement":"Text in the output PDF was selectable and matched the source document exactly."},{"tag":"ACCESS","statement":"The searchable PDF tool requires no account, no email address and no payment card."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages when building the text layer."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool PDFs one hour after conversion."}]},{"url":"https://visionparse.app/tools/screenshot-to-text","name":"Screenshot to Text","lastUpdated":"2026-08-01","measured":{"claim":"A screenshot-style image read in 0.39 s, against 0.52 s for the same content as a scanned page.","value":"0.39 s","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"Dark-mode screenshots were not part of this measurement, and light-on-dark text is a known weak case. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"SPEED","statement":"A screenshot read in 0.39 seconds, the fastest of every input format we measured."},{"tag":"ACCURACY","statement":"Screen-rendered text needs no deskewing, so it reads faster than a photographed page."},{"tag":"LIMIT","statement":"Interface text below 400 pixels of image width dropped to 56.8% accuracy in our sweep."},{"tag":"ACCESS","statement":"The screenshot to text tool requires no account, no email address and no payment card."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in screenshots, including Arabic, Sinhala and Tamil."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool screenshots one hour after conversion."}]},{"url":"https://visionparse.app/tools/table-to-csv","name":"Extract a Table from a PDF to CSV","lastUpdated":"2026-08-01","measured":{"claim":"Rows are recovered reliably; column boundaries are not. Three rows correct, four positional clusters for three columns.","value":"3 of 3 rows","measuredOn":"2026-08-01","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"One bordered table is a narrow sample. Borderless tables, merged cells and multi-line cells are materially harder and are not represented here. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"SPEED","statement":"VisionParse extracted a three-column table in 0.27 seconds on an 8-vCPU server."},{"tag":"ACCURACY","statement":"Row detection recovered all three rows of the test table with every cell read correctly."},{"tag":"LIMIT","statement":"Column clustering produced four groups for a three-column table, so column splits need checking."},{"tag":"ACCESS","statement":"The table to CSV tool requires no account, no email address and no payment card."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages inside tables, including Arabic, Sinhala and Tamil."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool files one hour after conversion."}]},{"url":"https://visionparse.app/tools/jpeg-to-text","name":"JPEG to Text","lastUpdated":"2026-08-02","measured":{"claim":"The JPEG encoding variant does not change recognition: baseline, progressive and 4:4:4 all read at 100%.","value":"100% on all 3 encodings","measuredOn":"2026-08-02","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"High-contrast printed text — the friendliest case for JPEG. Low-light photographs may respond differently to chroma subsampling. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"Baseline, progressive and 4:4:4 JPEG encodings of the same document each read at 100% character accuracy in our measurement."},{"tag":"SPEED","statement":"VisionParse read a JPEG invoice in 0.61 seconds on an 8-vCPU server."},{"tag":"LIMIT","statement":"A quality-10 JPEG — 27.9 KB against 95.5 KB at quality 85 — still read at 100% on printed text."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in JPEG files, including Arabic, Sinhala and Tamil."},{"tag":"ACCESS","statement":"The JPEG to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool JPEG files one hour after conversion."}]},{"url":"https://visionparse.app/tools/heic-to-text","name":"HEIC to Text","lastUpdated":"2026-08-02","measured":{"claim":"iPhone HEIC reads directly at full accuracy: quality 85 and a 3.6× smaller quality 30 both scored 100%.","value":"100% at q85 and q30","measuredOn":"2026-08-02","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"Synthetic render, not a handheld photo — real iPhone shots add focus, shadow and perspective effects that are about the camera, not the format. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"HEIC at quality 85 and quality 30 both read at 100% character accuracy in our measurement."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in HEIC photos, decoded on our servers with no JPG conversion first."},{"tag":"SPEED","statement":"A HEIC page read in 1.0–1.2 seconds — the HEVC decode adds about half a second over JPEG."},{"tag":"LIMIT","statement":"HEIC holds the same photo in roughly half the bytes of JPEG, so the 3 MB free limit goes further with iPhone photos."},{"tag":"ACCESS","statement":"The HEIC to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool HEIC files one hour after conversion."}]},{"url":"https://visionparse.app/tools/tiff-to-text","name":"TIFF to Text","lastUpdated":"2026-08-02","measured":{"claim":"Fax-grade Group 4 compression costs nothing: 6.4 KB and 940.9 KB versions of the same page both read at 100%.","value":"100% at 6.4 KB (G4 fax)","measuredOn":"2026-08-02","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"Group 4 is 1-bit black-and-white: ideal for printed text, destructive for photographs and grey stamps. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"A Group 4 fax-compressed TIFF — 147× smaller than the uncompressed file — read at the same 100% accuracy in our measurement."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in TIFF files, and reads multi-page TIFFs in one pass."},{"tag":"SPEED","statement":"VisionParse read a TIFF invoice in 0.51 seconds on an 8-vCPU server."},{"tag":"LIMIT","statement":"An uncompressed A4 TIFF runs near 1 MB a page, so the 3 MB free limit fits about 3 uncompressed pages — LZW fits dozens."},{"tag":"ACCESS","statement":"The TIFF to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool TIFF files one hour after conversion."}]},{"url":"https://visionparse.app/tools/bmp-to-text","name":"BMP to Text","lastUpdated":"2026-08-02","measured":{"claim":"BMP file size is padding, not information: 2,822 KB and 103 KB versions of the same page both read at 100%.","value":"100% at 27× size difference","measuredOn":"2026-08-02","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"8-bit palette conversion can hurt continuous-tone photographs even though it did not hurt printed text. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"A 2,822 KB 24-bit BMP and its 103 KB 8-bit version read at identical 100% accuracy in our measurement — 27× the bytes bought nothing."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in BMP files, including RLE-compressed BMPs many tools fail on."},{"tag":"SPEED","statement":"VisionParse read a BMP invoice in 0.57 seconds on an 8-vCPU server."},{"tag":"LIMIT","statement":"A single 24-bit A4 BMP nearly fills the 3 MB free limit — convert to PNG first if you need to send more than one page."},{"tag":"ACCESS","statement":"The BMP to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool BMP files one hour after conversion."}]},{"url":"https://visionparse.app/tools/webp-to-text","name":"WEBP to Text","lastUpdated":"2026-08-02","measured":{"claim":"For document images, lossless WebP beats lossy on size and ties it on accuracy: 20.6 KB lossless vs 46.8 KB at quality 80, both 100%.","value":"100% · lossless smaller than q80","measuredOn":"2026-08-02","hardware":"Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4","limitation":"True for flat, high-contrast document images; photographs invert this — lossy wins on size for continuous tones. A single synthetic document per variant is a small sample and is not a substitute for a corpus. These figures describe format behaviour, not general accuracy. A 500-document corpus run is scheduled and will supersede them.","methodology":"https://visionparse.app/benchmark#methodology"},"facts":[{"tag":"ACCURACY","statement":"Lossless, quality-80, quality-30 and quality-5 WebP each read at 100% character accuracy in our measurement."},{"tag":"LANGUAGES","statement":"VisionParse recognises 161 languages in WEBP files, lossy and lossless alike."},{"tag":"SPEED","statement":"VisionParse read a WEBP invoice in 0.58 seconds on an 8-vCPU server."},{"tag":"LIMIT","statement":"Even quality-5 WebP read at 100% on printed text — resolution, not compression, is what breaks recognition."},{"tag":"ACCESS","statement":"The WEBP to text tool requires no account, no email address and no payment card."},{"tag":"RETENTION","statement":"VisionParse deletes free-tool WEBP files one hour after conversion."}]}]}