OCR, but backwards.
Make your documents worse.

Generate photorealistic smartphone photos of any document — with verified ground truth for every field. For OCR training, vision benchmarks, and agentic pipeline testing.

Try the demo

Open source·Ground truth verified·API · CLI · MCP

InputOutput

Same data. Drag to compare.

Interactive Demo

Try the pipeline

Configure your input, select variations, and watch penquify run — step by step.

Photo Variations (2 selected)

Using cached example — modify the input to call the live API

🔬

Select your input, pick presets, and hit Run

2 free runs · No sign-up required

One command

From zero to verified dataset in 30 seconds

~ penquify

# Install

pip install penquify && playwright install chromium

# Full demo: PDF + 8 verified photo variations

penquify demo

# Upload any PDF, generate realistic photos

penquify photos --image invoice.pdf --presets blurry coffee_stain

# Describe what you want in plain English

penquify config --text "folded paper with grease, shot on old Motorola"

# Batch: 50 documents × 4 variations = 200 verified images

penquify dataset --count 50 --presets full_picture blurry coffee_stain cropped_header

“I needed to test my vision pipeline but I had 12 real documents. I needed 1,200 variations — blurry, folded, stained, cropped — with known ground truth for every field. I also needed the agent to do lookups after extraction, so the data in the photo had to be real and verifiable.”

The pipeline

From structured data to verified synthetic photos in one call. Every generated image has a ground-truth manifest.

Input

APIUI

JSON schema, uploaded PDF, or raw image. Define the data you need in the output: OC numbers, item names, quantities, dates. Or upload an existing document and we detect the schema.

Render clean document

Jinja2 HTML templates produce a pixel-perfect PDF. Dispatch guides, invoices, POs, BOLs — or bring your own template. Every field maps to a schema key.

Apply photo variations

Gemini generates a realistic photo using your variation config. Camera model, paper deformation, stains, blur, angle — every variable is controllable or randomizable.

Verify ground truth

Key

A second model pass reads the generated photo back and checks every schema field against the source data. If a field is wrong, it gets flagged and re-generated. Output: a verified image + ground truth JSON.

Occlusion manifest

Key

If a variation intentionally occludes data (crop, stain, fold), the manifest reports exactly which schema fields are affected and why. Your benchmark knows what should fail.

Output dataset

Clean PDF + N verified photos + ground truth JSON + occlusion manifest per image. Ready for training, benchmarking, or feeding into your agentic pipeline.

Same document. Different nightmares.

Every photo generated from the same clean PDF. Each preset targets a different real-world failure mode.

full_picture

Clean handheld, 90% frame

folded_skewed

Dog-ear, crease, 6deg tilt

zoomed_detail

Close-up OCR, oblique 25-30deg

blurry

Motion blur, partial legibility

coffee_stain

Stain over text, partial obstruction

cropped_header

Top 10-15% cut off

Case Study

Automating receipt with only purchase orders

A real scenario where penquify was built out of necessity. We had POs in the ERP. We had zero dispatch guides. We needed hundreds of realistic test photos.

The problem

Extract data from photos of dispatch guides taken by warehouse workers
Match items against purchase orders in the ERP
Handle name mismatches — supplier says “PAPA PREFRITA CONGELADA”, ERP says “PAPAS FRITAS 10MM”
Handle unit mismatches — guide says 12 CJ, PO says 150 KG, no weight-per-case given
Handle quantity discrepancies at two levels: guide vs PO, then physical count vs guide
Create ERP material documents with the correct batch, matching PO position

We had POs. We had zero dispatch guides. We needed hundreds of realistic test photos to validate the pipeline end-to-end.

What penquify did

Generated dispatch guides from PO data

Read the PO from the ERP (items, quantities, supplier). Generated a realistic dispatch guide PDF with different names — the way a real supplier writes them, not ERP master data names.

Introduced realistic discrepancies

Some items with different units (CJ vs KG, UN vs KG — no conversion on the guide). Some with 5% less quantity. Some with supplier-specific jargon.

Generated warehouse photos

6 variations per document: clean handheld, folded, blurry, coffee-stained, cropped header, zoomed detail. Each looks like a warehouse worker snapped it at a loading dock.

Verified every field

Ground truth JSON for each photo. Occlusion manifest saying exactly which fields are hidden in the coffee-stained or cropped versions.

Fed to the agent pipeline

Photo → vision extraction → PO matching → discrepancy detection → ERP write. Tested 100+ scenarios in hours instead of months of real operations.

The kind of mismatches real documents have

Penquify generates these programmatically. The ground truth knows the correct mapping.

Dispatch Guide (supplier)	Purchase Order (ERP)	Challenge
PAPA PREFRITA CONGELADA CORTE GRUESO 12 CJ	PAPAS FRITAS 10MM 150 KG	Different name + different unit. Guide doesn't say weight per case. Agent must ask.
MOZZARELLA RALLADA PREMIUM 115 KG	QUESO MOZZARELLA RALLADO 120 KG	Different name + 5kg short. Agent must flag discrepancy.
JENGIBRE FRESCO PELADO 2 UN	JENGIBRE 0.5 KG	Unit mismatch (UN vs KG). No weight per unit on guide. Agent must ask.
LIMON SUTIL FRESCO 24 KG	LIMON SUTIL 25 L	KG vs L + 1 unit short. Two problems at once.
MENTA FRESCA ATADO 10 UN	MENTA FRESCA 2 KG	Atados vs KG. No weight per atado. Agent must ask.

Everything is configurable

Camera model is free text. So is the angle. And the stain type. And the background. Every field is a knob.

⌁

22 camera presets + free text

Samsung Galaxy S7 through iPhone 14, budget Androids, rugged field devices. Or write anything: "Nokia 3310 with cracked screen"

◫

Paper physics

Curvature, folds (dog-ear, middle vertical, multiple), wrinkles, corner bends, edge curl. Paper behaves like paper.

◉

Damage & contamination

Coffee stains, water damage, grease marks, ink smudges, torn edges, dirt. Each with type, location, opacity, and text obstruction level.

⊿

Geometric distortion

Perspective skew, keystone distortion, rotation (0-15 deg), oblique angles up to 45 deg. Simulate every handheld capture angle.

◐

Lighting & artifacts

Glare hotspots, uneven lighting, shadow bands, finger shadows, JPEG compression (none to heavy), motion blur with direction control.

⊞

Natural language config

"Blurry photo with coffee stain, strong angle, old Samsung, paper folded in half" — AI converts your description to variation JSON.

✓

Ground truth verification

Every generated image is read back and verified against the source schema. Garbled fields get flagged and regenerated. You get verified data.

◎

Occlusion tracking

When a crop, stain, or fold hides data, the manifest says exactly which fields are affected and why. Your benchmark knows what should fail.

✋

Hand & context

Hand visible or not, grip type (thumb corner, pinched edge, both hands), gloves (warehouse, latex, none). Background as blurred context.

Every entry point

Start from wherever you are in your pipeline.

From scratch

JSON Schema → Photos

Send document data as JSON. Get back PDF + verified photos with ground truth.

Existing document

Upload PDF → Variations

Upload any PDF or image. We detect the schema, then generate N realistic variations with known ground truth.

Reference photo

Seed Image → Style Transfer

Provide a reference photo (lighting, background, camera style). New documents match that visual style.

Batch

N Docs × M Variations

Generate thousands of document-variation combinations. Progress tracking, S3 upload, parallel generation.

Fine-tuning

Training Pairs

Output image + ground truth pairs in formats ready for fine-tuning vision models: JSONL, COCO, or custom schema.

Pipeline testing

E2E Agent Tests

Generate documents with known data, feed them to your agent, verify extractions + downstream lookups match ground truth.

Built for real problems

Not another demo tool. Built because we needed it.

OCR Training

Generate 10,000 training pairs

N documents × M variations with ground truth labels. Export as JSONL for fine-tuning. Systematically vary difficulty: clean → blurry → stained → cropped.

Vision Benchmark

Measure extraction accuracy

Run your OCR/vision model on a controlled dataset. Know exactly which fields should be extractable vs occluded. Compare models objectively.

Agentic Testing

End-to-end agent validation

Generate documents with known data. Feed photos to your agent. Verify it extracts correctly AND does the right downstream actions.

Cold Start

Not enough real documents

You have 12 real invoices but need 1,200 test cases. Generate synthetic variations that cover every edge case your production pipeline will see.

Every interface

Terminal

CLI

penquify demo

Import

Python

from penquify import ...

HTTP

REST API

FastAPI + Swagger

AI Agent

MCP Server

5 tools for Claude

Plugin

Agent SDK

Drop-in tool list

Cloud

penquify.com

UI + API credits

Pricing that scales with you

Self-host free forever. Cloud for convenience. Enterprise for compliance.

Open Source

$0forever

Self-hosted. Bring your own Gemini key.

All document templates
All photo presets + custom
CLI + Python library
REST API (self-hosted)
MCP server + Agent SDK
Ground truth verification
Community support

Need more credits?

Buy additional credit packs anytime. Credits never expire and stack with your subscription.

MIT Licensed·No credit card required for trial·BYOK on all tiers·Cancel anytime

Ready to build better vision AI?

Start generating verified document datasets in 30 seconds. Open source, or let us handle the infra.

Get Started Free Talk to Us

OCR, but backwards.Make your documents worse.