Pipelines
A pipeline is a named analysis configuration you pass in the pipeline field of every verification request. A pipeline can chain detection models, inspect provenance data, or combine several analysis steps before returning its results. You choose the pipeline; it orchestrates the models, rules, routing, and post-processing for you.
Most pipelines analyze the media itself to ask whether it was generated or altered by AI. The c2pa_analysis pipeline instead reads the image's declared Content Credentials to assess provenance trust and semantic reliability. Pipelines also differ in the media they accept, the depth of analysis they perform, and the artifacts they return. See the Verification API Deep Dive for how a pipeline fits into the request lifecycle.
Available pipelines
| Pipeline | Status | Media | Description |
|---|---|---|---|
| cantaloupe | Public | Images | Deepfake image detector combining the Cantaloupe and Azalea models |
| honeydew | Closed beta | Images | Source attribution and tampering localization combining the Azalea and Honeydew models, with downloadable localization artifacts |
| pitaya | Closed beta | Videos | Detects whether a video is fully AI-generated or authentic via spatio-temporal analysis (Pitaya model) |
| c2pa_analysis | Closed beta | Images | Validates C2PA provenance and evaluates the semantic impact declared by an image's Content Credentials |
Additional pipelines are in development and available to select customers. Contact Truebees for early access.
How each pipeline works
cantaloupe — the standard authenticity check
The cantaloupe pipeline is the everyday image verification path. It answers the most common question: is this a real photograph, or was it AI-generated? It returns a binary verdict (AI_GENERATED / NOT_AI_GENERATED) and a confidence_score, so you get both the result and how certain the system is.
It runs two models in sequence:
- Azalea first classifies the image's social-media origin — whether it has been shared or processed through a platform such as Facebook or Instagram. This matters because platforms re-encode images on upload, and that re-compression introduces artifacts that can otherwise be mistaken for signs of manipulation. Azalea detects these traces and routes the image to the appropriate inference weights.
- Cantaloupe then performs the actual deepfake detection on the correctly-routed image, producing the final AI-generation verdict.
This social-media awareness is what keeps the verdict reliable on real-world images that have passed through messaging apps and social feeds.
Choose cantaloupe when you need a fast, dependable real-or-fake answer at scale. It is publicly available to any account with an active subscription.
Read the full cantaloupe page →
honeydew — source attribution and tampering localization
The honeydew pipeline performs a deeper analysis. Rather than a binary verdict, it attributes the image's source across several categories — authentic, GAN-generated, diffusion-generated, AI-partially-generated, or human-partially-edited — and it localizes which regions of the image appear manipulated.
It also runs two models in sequence:
- Azalea again determines social-media origin and routes between social and non-social inference weights, for the same reasons as in cantaloupe.
- Honeydew then classifies the image origin across the categories above and produces a per-category probability breakdown alongside the headline verdict.
On top of the verdict, Honeydew emits two downloadable artifacts:
| Artifact | What it is |
|---|---|
localization_map | A heatmap highlighting manipulated regions (red) versus authentic regions (green) |
localization_overlay | That heatmap blended over the original image for easier reading |
Artifacts are retained for 24 hours after the verification completes and are fetched by id from the artifact endpoint. They give you visual, defensible evidence — useful when you need to explain or justify a decision rather than record a bare yes/no.
Choose honeydew when you need to understand and demonstrate how an image was produced and where it was altered — for investigations, content-moderation review, or any case where a per-category breakdown and visual evidence add value. It is currently in closed beta; contact support@truebees.eu to request access.
pitaya — video deepfake detection
The pitaya pipeline is the first Truebees pipeline for video. It answers the same core question as cantaloupe — is this AI-generated, or authentic? — but for footage rather than still images, returning a binary verdict (AI_GENERATED / NOT_AI_GENERATED) and a confidence_score.
Unlike the image pipelines, it runs a single model:
- Pitaya is built on a foundational vision transformer backbone and analyzes spatio-temporal artifacts — the inconsistencies a generator leaves both within a frame and across consecutive frames. It samples temporally distributed clips across the whole video and extracts contiguous patches from consecutive frames, then aggregates the clip-level signals into one verdict.
Choose pitaya when you need to verify whether a video is AI-generated. It targets fully generated or fully authentic footage, not partially edited clips. It is currently in closed beta; contact support@truebees.eu to request access.
c2pa_analysis — provenance trust and semantic reliability
The c2pa_analysis pipeline reads an image's C2PA Content Credentials rather than inspecting its pixels. It returns two independent conclusions:
- Provenance validation reports whether a manifest is present, whether its signature and hashes verify, who signed the Active Manifest, and whether that signer is recognised by the configured trust list.
- Semantic reliability follows the declared history from the Active Manifest to the First Manifest and classifies the recorded actions by their impact on the image's fidelity as a depiction of physical reality.
Its semantic outcome is RELIABILITY_HIGH, RELIABILITY_MEDIUM, RELIABILITY_LOW, or RELIABILITY_ABSTAIN. Provenance validity and trust are returned separately in the evaluation metadata; neither changes the semantic outcome. The pipeline also reports declared and suggestive evidence of AI involvement found in the manifest.
Choose c2pa_analysis when you need to inspect an image's declared origin and editing history, verify who signed that record, or distinguish trustworthy provenance from what the record says about the content. It is currently in closed beta; contact support@truebees.eu to request access.
Read the full c2pa_analysis page →
Which pipeline should I choose?
| If you need… | Use |
|---|---|
| A fast binary real-or-fake verdict with a confidence score, at scale | cantaloupe |
| Source attribution across categories plus downloadable tampering-localization artifacts | honeydew |
| A real-or-fake verdict on a video | pitaya |
| C2PA manifest validation, signer trust, and declared semantic history | c2pa_analysis |
Not sure which fits your use case? Contact Truebees and we'll help you choose.