AI models for image, video and audio comparison matrix 2026

The most important technical terms from this comparison, explained neutrally.

Content Credentials

Adobe's implementation of the C2PA standard: visible provenance labels on images and videos, who created and edited what, with which tool.

Copyright Shield

OpenAI's commitment to defend business and API customers against copyright claims arising from generated content and to cover the costs.

IP indemnity

The vendor indemnifies you against third-party claims if generated content infringes someone else's rights, a key procurement criterion for media AI.

voice cloning

Replicating a real person's voice from short audio samples. Legally sensitive: consent and EU AI Act transparency duties apply.

Open-Weights

The model weights are freely available, the model can run on your own hardware. Not necessarily open source in the strict sense; check the licence in detail.

indemnity

The vendor assumes liability and defence costs if third-party claims arise from generated content.

watermark

Watermarks in AI content, visible or invisible (e.g. SynthID). Helps with labelling duties; depending on the vendor it may not be switchable.

SynthID

Google's invisible, robust watermark for AI-generated images, video and audio, always on for Google models and machine-verifiable.

AI Act

EU AI regulation: governs transparency and labelling duties for AI-generated content and deepfakes. Effectively relevant for Swiss vendors with EU customers too.

GDPR

EU General Data Protection Regulation. Relevant for Swiss companies alongside the FADP as soon as data of EU persons is processed.

C2PA

Open industry standard (Adobe, Microsoft, Google et al.) for cryptographically signed provenance data in media files, the technical basis for provenance duties.

GGUF

Compressed format for quantised AI models, cuts memory needs substantially and lets large models run on smaller hardware.

ARR

Annual recurring revenue, some «free» licences only apply below an ARR threshold.

self-host

Running the software/model on your own infrastructure instead of the vendor's, full data control, full operational burden.

TTS

Text-to-speech synthesis. Enterprise criteria: voice quality, latency, languages, cloning controls.

US Cloud Act

US law that can oblige American providers to hand data to US authorities, even when servers are located in Europe. Relevant for Swiss data-protection assessments.

MoE

Mixture of Experts, a model architecture where only part of the model is active per request. High capability at lower running cost.

Proprietary

Closed licence: usage only via the provider, no insight into model or code, no self-hosting.

Apache 2.0

Permissive open-source licence including commercial use and a patent-protection clause.

CN origin

Chinese vendor: uncritical when self-hosting the open weights; Chinese data law applies when using the vendor's API.

Hybrid

Flexible deployment: cloud, your own private cloud, or fully in your own data centre.