Image engine

The Image Engine verifies claims about images using deterministic methods first, with VLM fallback only for complex semantic claims.

Features

Metadata extraction — dimensions and format (PNG, JPEG, GIF, WebP)
Size verification — exact dimension comparison from metadata
Claim classification — routes claims to the appropriate verifier
Advisory VLM cross-check — optional model fallback for semantic claims

Usage

from qwed_sdk import QWEDClient

client = QWEDClient(api_key="qwed_...")

# Verify claim
result = client.verify_image(
    image_path="chart.png",
    claim="The image is 800x600 pixels"
)

print(result.status)      # "VERIFIED" (deterministic metadata proof)
print(result.is_verified) # True
print(result.result["developer_fields"]["analysis"]["metadata"])  # {"width": 800, "height": 600, "format": "png"}

Claim types

Claim Type	Method	Deterministic?
Size/Dimensions	Metadata parsing	✅ 100%
Format	Header detection	✅ 100%
Color	Pixel sampling	⚠️ Partial
Text	OCR required	❌ VLM
Semantic	Understanding	❌ VLM

Result contract

The Image Engine returns a DiagnosticResult whose status depends on which path handled the claim:

Path	`DiagnosticResult.status`	Notes
Deterministic evidence supports the claim	`VERIFIED`	Emits a `proof_ref` bound to pixel or metadata evidence.
Deterministic evidence refutes the claim	`BLOCKED`	Fail-closed refutation.
VLM cross-check only	`UNVERIFIABLE`	VLM results are recorded in `advisory_checks`; the engine never emits `VERIFIED` from a VLM alone.

This mirrors the 3-tier engine classification: a VLM signal is advisory and cannot promote a claim to VERIFIED.

VLM cross-check (advisory)

Semantic claims (color, text, visual understanding) cannot be proven deterministically. The core ImageVerifier supports a VLM fallback for these claims when instantiated with use_vlm_fallback=True; the VLM output is recorded in advisory_checks and does not change the diagnostic status. The public /verify/image endpoint runs with use_vlm_fallback=False, so via the SDK a semantic claim resolves to UNVERIFIABLE with the VLM path listed as advisory-only:

result = client.verify_image(
    image_path="photo.png",
    claim="The person is smiling"
)

print(result.status)      # "UNVERIFIABLE" (deterministic path cannot verify)
print(result.result["developer_fields"].get("advisory_checks", []))  # [] for SDK/API path

Supported providers for image verification

The Image Engine supports multimodal verification through any provider that accepts image inputs. When you set ACTIVE_PROVIDER=gemini, QWED uses Google Gemini’s native multimodal capabilities for semantic image claims. Supported image formats are JPEG, PNG, and WebP.

export ACTIVE_PROVIDER=gemini
export GOOGLE_API_KEY=your-google-api-key

See LLM configuration for full setup instructions.

Overview

Core Engines

Analysis Engines

Specialized Engines

Features

Usage

Claim types

Result contract

VLM cross-check (advisory)

Supported providers for image verification

​Features

​Usage

​Claim types

​Result contract

​VLM cross-check (advisory)

​Supported providers for image verification

Features

Usage

Claim types

Result contract

VLM cross-check (advisory)

Supported providers for image verification