Provenance-first restoration · Backblaze B2 + Genblaze

AI restoration invents history. Trueprint proves what's real.

Restore old photos and audio like the best AI tools, then get a cryptographically verifiable record of exactly which pixels are original, which are enhanced, and which the AI fabricated.

Try the demo Verify a file Watch 3-min demo

Backblaze B2 Genblaze SDK C2PA Content Credentials EU AI Act · Art. 50 OAIS archive model

Live sample

Real restored archival sample from Backblaze B2

Original Enhanced Fabricated

RESTORED · loading a real sample from Backblaze B2…

A 2026 Library of Congress call-to-action urges libraries, archives, and museums to keep AI-affected collections authentic, transparent, and verifiable, from creation through access. Trueprint is built for exactly that.

Paraphrased from “Content Authenticity & Provenance in the Age of AI: A Call-to-Action for the LAMs Community,” The Signal, Library of Congress, 2026 · source

The problem

Restoration doesn't recover the past. It guesses it, silently.

When an AI colorizes a 1920s photograph, it isn't recovering the true color of a dress or a storefront sign. It's inventing a beautiful, plausible fiction. The moment that version circulates, the historical record is quietly corrupted, and no one can tell which parts were real.

Invented, not recovered

Colorization is a guess

A black-and-white master contains no color data. Every hue an AI adds is inferred from what "looks right", not from the original scene.

Detail from nowhere

Upscaling hallucinates

Super-resolution and face restoration fabricate fine detail that never existed in the source, rebuilding faces from a generic prior.

No paper trail

Consumer tools are provenance-blind

Remini, MyHeritage, Palette.fm and others give you a pretty result with zero record of what changed. Institutions can't cite what they can't verify.

The core idea

Every pixel gets a verdict: original, enhanced, or fabricated.

Trueprint classifies the restored image at the region level, so trust is never all-or-nothing. This three-class model is the language of the whole product.

Original

Real signal, preserved

Pixels carried through from the master untouched, within tolerance. This is genuine historical data, the part you can cite without an asterisk.

Enhanced

Real signal, cleaned

Original data that was denoised, de-scratched, or contrast-corrected. Nothing invented, the underlying signal is real, just clarified.

Fabricated

Invented by AI

Content with no basis in the master: inpainted damage, all colorized chroma, hallucinated upscale detail. Beautiful, but a guess.

How it works · the pipeline

One orchestrated pipeline, many specialist models.

Genblaze chains best-in-class providers into a single restoration run. Each step declares fallback models, so a provider outage degrades gracefully instead of failing, and every step is recorded in a hash-verified lineage.

STEP 01

Analyze & detect damage

A vision model reads the scan: describes the scene, finds scratches, tears and holes, and emits a damage mask that drives the repair.

STEP 02

Repair & inpaint

An image-to-image model fills damaged regions. The damage mask fed in becomes the fabrication record out, provenance captured at the source.

STEP 03

Detect faces

When faces are present, YuNet flags each one and reports how confident the AI's colour is inside it. We do not rebuild faces — that would invent identity — we mark the region viewers scrutinise most.

STEP 04

Colorize ×2 · corroborate

Two independent providers (Google Gemini + OpenAI gpt-image) colorize the same image. Where they agree, confidence is high; where they diverge, we flag the region as a guess.

STEP 05

Upscale

Super-resolution recovers apparent sharpness. New high-frequency detail is measured against an honest upscale and marked as invented.

STEP 06

Restore audio

A subtractive spectral-gate pass cleans hiss and crackle — it can only remove energy, never invent audio (0% fabricated). Live in the app's Audio tab.

Verification loop (LLM-as-judge)

After restoration, a vision model compares the result to the master and flags historically implausible inventions, an added person, altered sign text. The verdict, score, and feedback are recorded in the manifest, so an implausible restoration is caught and disclosed, not silently shipped.

Hash-verified step lineage

Every step records its inputs, model, and output hashes in the manifest, so the whole restoration is an auditable record on Backblaze B2, not a black box.

The authenticity map · our technical core

Recorded at generation time, not reverse-engineered after.

Because Trueprint orchestrates every step, it knows exactly what each model changed and where. That's why the map is real and reproducible, not a plausible-looking overlay. Here's how each operation is classified, honestly.

Operation	What's fabricated	How we know & mark it	Class
Damage inpainting	Content filling holes, scratches, tears	The damage mask fed into the step is the fabricated region, recorded directly.	FABRICATED
Colorization	All chroma, color is a guess by definition	A grayscale master has no color, so 100% of hue is inferred. Disagreement between the two providers grades confidence.	FABRICATED
Super-resolution	High-frequency detail absent in the master	High-pass diff vs. an honest Lanczos upscale isolates the detail the model invented.	FABRICATED
Face region (detected)	Nothing — we do not touch identity	YuNet detects each face and reports the AI's colour confidence inside it; the box flags the most-scrutinised region.	FLAGGED
Audio denoise	Nothing — subtractive by construction	Spectral gating only removes energy (mask in [0,1]); it cannot synthesize audio, so fabrication is 0.	ENHANCED
Untouched regions	Nothing	Pixels identical to the master within tolerance.	ORIGINAL

The novel move

Multi-provider confidence

We run colorization across two independent providers, Google Gemini and OpenAI’s gpt-image (orchestrated via Genblaze), and measure per-pixel variance. Agreement means the inference is well-grounded; disagreement means the AI is genuinely guessing, surfaced as a confidence heatmap no other restoration tool produces.

The output

A plain-language verdict

Every result carries a one-line disclosure: “Structure preserved; all colour is AI-inferred (~52% of the image), colorizer confidence 54%.” Human-readable for a museum label; machine-readable for compliance.

Provenance & storage · Backblaze B2

The record outlives the file, and can't be quietly rewritten.

Every restoration produces a hash-verified manifest, embedded into the media and archived durably on Backblaze B2 alongside an immutable original master. Trust that travels with the file.

manifest.json · embedded + B2

{
  "asset_id": "tp_0x8f21",
  "master": { "sha256": "a1b7…9f3c", "object_lock": true },
  "pipeline": [
    { "step":"inpaint", "model":"opencv-inpaint",
      "operation":"FABRICATED", "region":"mask_02.png" },
    { "step":"colorize", "providers":2,
      "confidence":0.71 }
  ],
  "authenticity": { "original":0.55,
    "enhanced":0.07, "fabricated":0.38 },
  "disclosure": "~38% AI-inferred; all color; 1 face."
}

Dual archive on B2 (OAIS model)

Immutable masters + restored derivatives + manifests, in a hierarchical, queryable key layout, the system of record, not a dumb blob store.

Object Lock on masters

Backblaze B2 Object Lock makes originals provably unalterable, giving the archive tamper-evident, retention-grade integrity.

Off-platform verification

The Verify page reads the manifest embedded in any downloaded Trueprint file and re-checks its hash against B2, a green pass, or a red tamper alert.

C2PA Content Credentials & Article 50 ready

Field names align with the C2PA content-provenance standard, and the disclosure doubles as EU AI Act Article 50 machine-readable AI labeling, enforced from Aug 2, 2026.

Open the Verify page

Who it's for

Anyone who has to trust a restored image, and be trusted with it.

Institutions

Archives & museums

Publish restored collections with an auditable line between conserved fact and artistic reconstruction.

Everyday

Genealogists & families

Bring heirloom photos back to life while keeping an honest note of what was invented, memory, not fabrication.

Media

Journalists & documentarians

Use restored archival footage on air with a provenance credit that survives fact-checking.

Legal

Courts & estates

Enter restored evidence with a chain of custody: what's original, what's model-generated, hash-sealed.

Built on

The stack that makes provenance possible.

Trueprint isn't a wrapper around one model, it's an orchestration and archival system. These pieces are load-bearing, not logos.

Storage · system of record

Backblaze B2

S3-compatible object storage holding the whole archive.

Immutable masters via Object Lock
Derivatives, manifests & step outputs
Queryable lineage catalog
Serve, verify & replay from storage

Orchestration · provenance

Genblaze SDK

Backblaze's open-source pipeline engine.

Multi-provider, multi-modal chaining
Fallback models & retries
Hash-verified embedded manifests
Parent to child run lineage

Models · inference

Providers & engines

What each step actually calls.

Vision LLM + LLM-as-judge — GMI Cloud (Gemini)
Colorize ×2 — Google Gemini + OpenAI gpt-image
Face detect (YuNet), super-res, audio denoise — local CV/DSP
Swappable per step, no re-architecting

Learn · the concepts in plain language

Everything you need to follow the whole idea.

New to provenance and content authenticity? These are the terms behind Trueprint, defined simply. Expand any one.

+ Provenance

The documented history of where something came from and everything that happened to it. For a restored image, provenance answers: what was the original, which models touched it, and what did each one change?

+ Authenticity map

A region-level overlay classifying every part of a restored image as original, enhanced, or fabricated, so you can see at a glance what's genuine and what the AI invented.

+ Fabricated vs. original

Original = real signal from the source. Fabricated = content invented by a model with no basis in the source (colorized hue, inpainted damage, hallucinated detail). Trueprint never blurs the two.

+ Manifest

A structured record, a small JSON document, listing every pipeline step, the models used, the regions they changed, and hashes of the input and output. Trueprint embeds it inside the media file and stores a copy on B2.

+ Hash / fingerprint

A SHA-256 hash is a short string uniquely derived from a file's exact bytes. Change one pixel and the hash changes completely, which is how tampering is detected during verification.

+ Multi-provider corroboration

Running the same task across two independent providers (Google Gemini + OpenAI gpt-image) and comparing results. Agreement signals a well-grounded inference; disagreement reveals genuine uncertainty, Trueprint turns this into a confidence score.

+ C2PA

The Coalition for Content Provenance and Authenticity standard, a cryptographically signed record of a file's origin and edit history, backed by Adobe, the BBC, the New York Times and others. Trueprint's manifest aligns with it for interoperability.

+ OAIS & Object Lock

OAIS is the reference model for long-term digital archiving (masters, derivatives, metadata kept distinct). Object Lock is a Backblaze B2 feature that makes a stored object immutable for a set period, so an original master can't be silently altered.

+ EU AI Act · Article 50

The EU AI Act's transparency rule, enforced from Aug 2, 2026, requiring AI-generated or manipulated media to carry machine-readable disclosure. Trueprint's provenance record satisfies this by design.

Restoration you can cite

See exactly what the AI invented.

Upload a damaged photo, watch it come back to life across multiple models, and get a verifiable record of every choice, stored durably on Backblaze B2.

Try the demo Verify a file View the repo