Tie a statement to visible evidence

Check what an image can actually support

Write one concrete statement, compare quoted wording with local OCR, and optionally ground the statement to pixel regions with a private browser vision model.

Ground a visual claim

Completely free No account or payment Image stays on this device

USE THIS WHENUse this when someone makes one concrete statement about wording, count, color, condition or an object visible in an image.SUPPORTED INPUTOne JPEG, PNG or WebP · 32 MP

What you will getA clear result before technical detail

01A supported, contradicted or insufficient visible-evidence result

02Numbered OCR and visual evidence regions

03A local JSON report with checks, regions and limitations

Private analysis labFiles stay inside this browser

1 image · 280-character claim · optional 350–700 MiB model

Choose the image behind the claimJPEG, PNG or WebP · 25 MiB · 32 MP claim-evidence limit

Selection is processed locally

Direct textQuoted wording compared with bounded OCR lines

Grounded pixelsClaim-linked regions instead of a free answer

Honest uncertaintyMissing evidence stays insufficient, not false

Session diagnosticsNo processing errors

Stored only in this browser tab. Image bytes, filenames and metadata are never included.

No tool error has been recorded in this tab.

No upload or account Originals remain unchangedReview methodology

Useful and careful

What is built into the result.

Important findings come first. Technical fields remain available without taking over the page.

Tests one observable statement instead of inventing a general verdict

Uses direct quoted-text matching before optional model work

Treats missing OCR or model output as insufficient—not automatically false

Understand the method

A useful result with its reasoning attached.

This page explains what the lab measures, how to interpret it and where human review remains essential.

Turn a broad statement into checks the pixels can answer

An image can show a printed serial number, a visible crack, a colored warning sign or a count of bounded objects. It usually cannot prove who caused an event, when it happened, whether a person is trustworthy or whether an unseen condition exists. The claim grounder begins by keeping that boundary visible. It accepts one short statement, removes control characters, separates quoted text, numbers, colors, negation and observable terms, then records that decomposition in the report.

Quoted wording receives the strongest deterministic check. If the user writes that a label says “SN-204,” the local OCR regions are compared with that exact normalized phrase. A matching bounded line can support the visible-text part of the claim without downloading a large vision model. Failure to read the phrase stays insufficient because low contrast, unusual lettering, rotation or a partial crop can make OCR miss text that a person can still see.

One claim at a time
Exact quoted-text comparison
Count, color and negation decomposition
No claim expansion beyond the image

Ground the statement to regions instead of trusting a fluent caption

The optional advanced pass sends the user's statement and local image object to a dedicated Florence-2 WebGPU worker. Phrase grounding returns bounded regions related to the claim, while a separate detailed-caption pass supplies context. The free caption never decides the result by itself. It can help a reviewer notice vocabulary overlap, but only direct OCR evidence or a bounded grounding region may support or conflict with the statement.

Overlapping grounding boxes are deduplicated before count comparison. An exact count match can support a concrete positive claim, but a different model count is not treated as an automatic contradiction because the model can miss an object or draw two boxes around one object. Negative claims are equally conservative: a returned bounded region can conflict with “there is no …,” while no returned region remains insufficient rather than proving absence.

Dedicated browser worker
Bounded phrase-grounding regions
Independent context caption
Conservative count and absence rules

Read supported, contradicted and insufficient precisely

Supported means that at least one direct text match or bounded region agrees with the observable statement inside the selected image. It does not authenticate the file, establish the truth of the surrounding story or prove anything outside the frame. Contradicted is reserved for a direct visible conflict, such as a claim denying quoted wording that the OCR pass found, or a negative statement paired with a returned claim region. Every conflict still requires human review.

Insufficient is a useful result, not a processing failure. It means the available evidence cannot justify support or contradiction. The evidence board shows each check, its outcome, the exact limited observation and any numbered region. The JSON export excludes the source image and filename while preserving the normalized claim, check results, boxes, model source, image-upload count and limitations so another reviewer can understand how the result was produced.

Supported is not real-world verification
Contradiction requires visible conflict
Insufficient preserves uncertainty
Image-free evidence JSON

Three simple steps

Know what happens before you start.

Use this when someone makes one concrete statement about wording, count, color, condition or an object visible in an image.

StateWrite one observable claim; put exact visible wording in quotation marks.
CompareCheck bounded OCR evidence first, then optionally ground the claim to pixels with local WebGPU.
InterpretReview supported, contradicted or insufficient with every check and limitation attached.

RESULT ORDER

01 · A supported, contradicted or insufficient visible-evidence result02 · Numbered OCR and visual evidence regions03 · A local JSON report with checks, regions and limitations04 · Limits and next step

Clear before you rely on it

Questions this tool should answer upfront.

Short answers keep important privacy, evidence and professional-use limits visible.

01Can this tool fact-check a news photo?

It can only test a concrete statement against text and regions visible in that image. It cannot verify the event, source, date, location or surrounding story, so it is not a complete news fact-checker.

02Why does a missing object return insufficient instead of contradicted?

Grounding models can miss small, obscured or unusual objects. Absence from model output is therefore not reliable evidence that the object is absent from the image.

03Does the image leave my device?

No image upload request exists. Local OCR and WebGPU inference use the browser-selected File. During development the optional model can come from Hugging Face; production is configured for same-origin model assets.

04What kind of claim works best?

Use a short observable statement such as a quoted label, a visible condition, an object count or a color. Avoid claims about motive, identity, time, ownership or facts outside the frame.

Method referencesONNX Community Florence-2 model card and tasks Microsoft Florence-2 research paper NIST AI Risk Management Framework

Continue when useful

Your next step, without starting over.

Move to another page only when its outcome matches what you need.

Inspect file evidenceKeep metadata and visible claim evidence separate.Check Content CredentialsReview signed provenance without treating it as a truth verdict.

Important limitations

The result covers only what can be observed in this image · Vision grounding can miss, duplicate or mislabel a region · Supported visible evidence does not verify the depicted event, identity, place or date

Full limitations