Open Benchmark Masters, File Inspector & Verified Badges
Inspect our measurement methods and recorded artifacts: download our exact test inputs and Ground-Truth reference masters across all 6 online utility & AI categories, drop any visual tool's output into our client-side inspector below for preview metrics, or query our JSON/CSV dataset (224 reports across 86 platforms).
Browser File Inspector & Published Report Matching
Run our canonical test input on any online utility or AI website, then drop the exported image or .docx file here to inspect file metrics. Exact matches link to published report scores; other files receive no inferred score or ranking:
1. Canonical Test Inputs & Ground-Truth Masters (All 6 Categories)
Download the exact test input files (images / PDF) and Ground-Truth reference masters used across all RealTest evaluations:
AI Background Remover: Chroma-Key Curly Hair Alpha Master
Constructed by extracting sub-pixel Ground-Truth alpha (alpha_gt.png) from studio chroma-key subject photography and compositing over a cluttered warm bookstore-café plate to measure mad_band, SAD(k), fg_recall, and bg_leak.
AI Image Generator: 5-Point Hard Constraint Prompt
Simultaneously tests: (1) Exact text rendering "REALTEST 2026" (25 pts); (2) Count & color attribute binding: 3 red apples + 2 blue ceramic mugs (20 pts); (3) 5-finger right-hand anatomy (20 pts).
A close-up studio photo of a vintage brass robot holding a wooden sign with the exact text "REALTEST 2026", with 3 red apples and 2 blue ceramic mugs on a dark oak table, showing the robot's right hand with 5 distinct fingers, 1:1 square 
AI Image Upscaler: 512px Degraded → 2048px Ground-Truth Master
Derived from a 2048×2048 Ground-Truth master with an embedded resolution calibration plate, degraded via 4× downsampling + σ=0.65 Gaussian blur + JPEG Q=82 into a 512×512 test input to measure PSNR, SSIM, edge_corr, and hf_ratio.
AI Object Remover: 3-Zone Synthetic Occlusion → 2048px Clean GT
Synthetically occludes a clean 2048×2048 bookstore-café Ground-Truth plate with 3 high-contrast signs (RESERVED, SALE 50%, FREE WIFI; baseline PSNR=6.15 dB) to measure masked restoration PSNR/SSIM and unmasked preservation PSNR.
Image Compressor: 4.14 MB High-Entropy Lossless PNG Master
Uses a 2048×2048 (4,237,168 bytes = 4,137.9 KB) lossless 24-bit PNG containing warm lighting gradients, fine book-spine typography, and dark shadow textures to measure file size reduction %, PSNR (dB), SSIM, and silent dimension downscaling.
PDF to Word: 2-Page Complex Table & Dual-Column PDF Master
A 2-page vector PDF master (document.pdf) featuring 4 headings, a 16-cell 4×4 financial/metrics table, two-column prose, and a Page-2 truncation token (REALTEST-PAGE2-VERIFIED-2026). Evaluated by parsing word/document.xml inside the exported .docx.
PDF Compressor: 1.66 MB 2-Page Vector Text & Image Stream Master
A 2-page 1,659.0 KB (1,698,823 bytes) PDF master containing 27 selectable vector text blocks, a 12-cell financial table, a Page-2 verification token, and high-res image streams. Stream inspection separates lossless vector-preserving image recompression from destructive full-page bitmap flattening.
Scanned PDF OCR: 644 KB 2-Page Raster Scan & Confusable Table Master
A 2-page raster-only scanned PDF master (659,826 bytes = 644.4 KB, zero embedded fonts and zero BT/ET operators, with 0.45° deskew tilt and scanner grain) containing white-on-navy header banners, a 12-cell confusable invoice table (B8-105S vs O0-I1l8), prose, and a Page-2 verification token.
PDF Merge & Split: 2-Document 3-Page Sequence, Table & Vector Stream Master
Consists of a 2-page primary PDF (part1_main.pdf, 268.0 KB) and a 1-page audit appendix (part2_appendix.pdf, 48.9 KB) totaling 3 pages (316.8 KB) with 3 page-sequence tokens (DOC1-P1 → DOC1-P2 → DOC2-P3), 12 cross-document table cells, and embedded image streams to verify page completeness, sequence fidelity, and vector text retention.
Photo Restoration: 1024px Controlled Blur, Crease Scratch & Film Grain Portrait Master
Constructed by degrading a clean 1024×1024 studio portrait Ground-Truth master (gt_1024.png) with optical defocus (σ=1.65), 2× down/upsampling, film grain, 5 synthetic vintage creases/scratches (scratch_mask.png, baseline scratch PSNR=6.02 dB), and JPEG Q=72 artifacts to measure global PSNR/SSIM, scratch-region repair PSNR, and facial ROI fidelity.
2. Embeddable assessment coverage badges (86 Platforms)
Product badges show the number of released task assessments, without a cross-task average or product certification. Choose task score badges on the public data page.
3. Public JSON API Schema & Citation Guide
GET https://realtest.org/api/benchmarks.json returns the full CC-BY-4.0 structured dataset updated on every build (no API key required):
{
"schema_version": "3.0",
"license": "CC-BY-4.0",
"citation_url": "https://realtest.org/methodology",
"total_platforms": 86,
"total_benchmarks": 224,
"benchmarks": [
{
"category": "image-compressor",
"rank": 1,
"domain": "squoosh.app",
"total_score": 96.6,
"honesty_score": 99,
"capability_score": 95.6,
"report_url": "https://realtest.org/report/squoosh.app",
"badge_svg_url": "https://realtest.org/badges/image-compressor/squoosh.app.svg"
}
]
}