TerraVis World-grounded visual consistency

Does your image obey the world?

Upload any generated image. TerraVis helps you identify 18 types of world‑consistency violations across objects, interactions, and scenes. You get an overall score and a breakdown of detected violations, with supporting evidence and severity levels.

Evaluate an image Browse examples
Stages
3
Violation probes
18
Score range
(0, 1]
Examples

01

Evaluate your image

Drop, paste or browse an image (JPEG, PNG or WebP, up to 15 MB).

Specimen

Sent to the judge only when you press Evaluate, and never stored.

Readout

IDLE
  1. 1Eligibility—
  2. 2Detection0/18
  3. 3Severity—
  4. SScore—
Major
—
Minor
—
Clear
—
Unparsed ?
—

Load an image to begin.

Live progress00:00.0

    Violation probes 18

    • Pending
    • Checking
    • Clear
    • Minor
    • Major
    • Unparsed

    No probe was run: the eligibility stage gave no score, so the 18 checks were skipped.

    Findings

    Examples Precomputed

    Scored offline with the same judge. Opening one replays its stored result.

      02

      How it works

      Three stages, each a small, checkable question to the judge. The score only counts what was actually found.

      1. Stage 1

        Eligibility

        Is this a single-frame, representational image of identifiable macroscopic objects or scenes? Photorealistic and stylized images both count.

        1 query

      2. Stage 2

        Violation detection

        18 yes/no probes run in parallel. Each “yes” names the affected aspect and one sentence of visible evidence.

        18 parallel queries

      3. Stage 3

        Severity

        Major if it affects the main subject, a foreground object or the central event; minor if it is background or peripheral.

        1 query per detection

      4. Score

        Aggregation

        S = exp(−λ ·  (Nmajor + α · Nminor))

        λ = 1.0α = 0.5

        range (0, 1] · 1 = no violations

      Score ladder

      Each major violation multiplies the score by e−λ, each minor one by e−λα. Every score corresponds to a count, so read results as counts.

      Nmajor 1
      Nminor 1

      = 0.223

        03

        The 18 violation types

        TerraVis breaks world consistency into three levels: 7 object-level, 8 interaction-level and 3 scene-level violation types. Each type is one probe in the detection stage.

        04

        Questions, privacy & limitations

        What does TerraVis measure?

        World-grounded visual consistency (world consistency for short): whether what the image depicts could exist in the physical world. That covers plausible object identity and structure, anatomy, surfaces and legible markings; physically sound contact, support, motion, media, optics, energy and thermal responses; and coherent relative scale, occlusion and region continuity. It evaluates both photorealistic and stylized images based on visual content alone. Characteristics inherent to stylized images, such as simplified textures, flat colors, and schematic lighting, are not considered violations in themselves.

        How do I read the score?

        S = exp(−λ·(Nmajor + α·Nminor)) with λ = 1.0 and α = 0.5. 1.000 means none of the 18 probes found a violation; one minor violation gives 0.607, one major (or two minor) 0.368, one major plus one minor 0.223. Each score maps back to violation counts, so compare results by those counts.

        Why did my image get “Not applicable”?

        TerraVis only scores single-frame representational images of macroscopic objects or scenes. The judge marks an image N/A when its primary content is one of:

          A physical object that happens to show text, a map or a menu (a sign in a street, a book on a table) is still eligible.

          What does “Undetermined” or “Unparsed” mean?

          Undetermined: the judge’s eligibility answer could not be parsed after 3 attempts, so no score is given. Unparsed on a single probe: that one answer could not be read and is counted as not detected.

          Is my image stored?

          No. Your image is held in memory only while it is queued and judged, and is never written to disk. The result, keyed by a hash of the file, stays in the server’s memory for about 3 minutes, so this page can reconnect if your connection drops and an identical re-upload gets the same result back; then it is discarded.

          Can TerraVis tell whether an image is AI-generated?

          No. TerraVis is a research metric for the physical and commonsense plausibility of an image, not a forensic or deepfake detector. A real photograph can receive findings, and a generated image can receive none.

          Limitations
          • The judge can be wrong: it can miss subtle violations and flag things that are fine. Read the evidence sentences, not just the number.
          • TerraVis is a research metric for world consistency. It is not a forensic or deepfake detector and says nothing about whether an image is real or generated.
          • Each probe reports at most one instance (the most salient), so several distortions of the same type count once.