-
-
Notifications
You must be signed in to change notification settings - Fork 0
Format Support
Jura Trace v1.0 is explicit about what it can and cannot meaningfully analyse. This page is the single source of truth for file format support, drawn from the April 2026 audit that compared the pipeline's actual detector paths against the formats marketed as supported.
If your file type is not listed under "Supported formats", it has no working detector path in this release. Jura Trace will refuse the file rather than return a trust score with no underlying signal. A score unsupported by analysis is worse than no score at all.
Protect (sign): Content Credentials (C2PA manifest), perceptual fingerprint (aHash, dHash, pHash), full metadata extraction.
Verify: Full forensic stack. ELA, JPEG Ghost (JPEG only), copy-move detection, the deepfake ensemble (GBM v4 + UnivFD v10onnx), CLIP Detection (optional), EXIF anomaly analysis, segmented ELA (Deep mode), colour temperature (Deep mode). On-demand tools (NPR, shadow consistency, splice boundary) available in Expert View.
JPEG, PNG, and TIFF are the Validator-Conformant signing formats. The C2PA Conformant Validator award (record ID 019d8d83-ed1c-787c-920c-8fad67b55cbe, awarded 6 May 2026) covers exactly these three formats plus WebP.
Protect (sign): Content Credentials and perceptual fingerprint for WebP. HEIC/HEIF and AVIF: the underlying c2pa-rs library can sign these formats, but they are outside the Validator-Conformant scope. Signing HEIC, HEIF, or AVIF files in v1.0 produces a valid C2PA manifest, but that manifest will not carry the Conformant Validator badge. These formats are therefore excluded from the Protect file picker in v1.0.
Verify: Deepfake ensemble, EXIF anomaly, CLIP Detection, segmented ELA. ELA and JPEG Ghost are codec-gated off for these formats because both techniques rely on JPEG discrete cosine transform block structure.
Note for iPhone users. iPhone photos arrive as HEIC by default. Jura Trace v1.0 on Linux requires the pillow-heif library to decode HEIC files; this is bundled in the v1.0 Linux installer.
The table below shows which automatic detectors run on each supported format. Detectors that are codec-gated off appear as ✗. Detectors that run on all supported formats appear as ✓. The three on-demand investigation tools (NPR, Shadow Consistency, Splice Boundary) are available for all supported formats regardless of mode; they are omitted from this table because they do not contribute to the trust score.
| Detector | JPEG | PNG | TIFF | WebP | HEIC/HEIF | AVIF |
|---|---|---|---|---|---|---|
| EXIF Anomaly | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| C2PA Provenance | ✓ | ✓ | ✓ | ✓ | ✗ | ✗ |
| Error Level Analysis (ELA) | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ |
| Noise Analysis | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Copy-Move Detection | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| AI Generation Detection (GBM + UnivFD) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| JPEG Ghost | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ |
| Segmented ELA (Deep mode only) | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ |
| Colour Temperature (Deep mode only) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| CLIP Detection (optional, Deep mode) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
Notes:
- ELA and JPEG Ghost rely on JPEG discrete cosine transform block structure and are not applicable to other codecs.
- C2PA signing is available for JPEG, PNG, TIFF, and WebP (Validator-Conformant scope). HEIC/HEIF and AVIF can be decoded for analysis but cannot be signed in v1.0.
- Segmented ELA is a regional variant of ELA and shares the same codec restriction.
- CLIP Detection requires the optional ~580 MB model download. The detector is skipped gracefully if the model is not installed.
These formats are excluded from the file picker because they have no working detector path in this release. The decision to exclude rather than return an unsubstantiated score was confirmed during user testing on 28 April 2026.
Not supported. Jura Trace v1.0 analyses still images only. Video files are excluded from both the Verify and Protect file pickers.
Video deepfake analysis (per-frame GBM/UnivFD ensemble plus temporal consistency signals) is designed and implemented in the codebase, but was deferred from v1.0 on 2 May 2026 pending calibration of the detection pipeline against Global Majority device capture (Tecno, Infinix, Samsung-A series, Xiaomi). Publishing a detector with known uncalibrated gaps on devices used by the majority of the world's journalists and researchers would not meet the project's honesty standard. Target release: v1.0.1.
Not supported. The audio deepfake model was deferred from v1.0 because the current training corpus covers only two speakers and one TTS engine. An honest detector requires retraining on ASVspoof and WaveFake with held-out speaker sets before publication. Target release: v1.0.1.
Not supported. PDF is excluded from the Verify file picker. Jura Trace v1.0 analyses raster still images only. The underlying c2pa-rs library does support PDF provenance, but the pixel-level forensic pipeline (ELA, noise analysis, deepfake detection) requires a raster image as input.
No detector path. Excluded from the file picker.
No Content Credentials, watermark, or fingerprint path for animated GIF. Excluded from the file picker for v1.0.
Out of scope for v1.0. No detector path.
A trust score that does not reflect a working detector is a worse outcome than no score at all. In testing before v1.0, a DOCX file dropped onto the pipeline returned 0.50 trust with zero underlying analysis. Users interpreted that as "clean" when the system had not checked anything. The same principle applies to audio and video: models trained on narrow corpora cannot defensibly assess voice notes or video clips submitted as evidence. It is better to refuse the file than to ship false confidence.
Last updated 2026-06-04 for Jura Trace v1.0.