Decision aid for marketing, communications & IT
AI models for image, video and audio – comparison matrix 2026
Generative media models are procurement-ready in 2026, but the risks differ from coding tools: output rights, provenance obligations and licence pitfalls decide, not benchmarks.
This matrix compares 32 image, video and audio models plus one model-agnostic workflow tool against the criteria that actually derail adoption in Switzerland and the EU: may the outputs be used commercially? Is there IP indemnification? Can the model run on-prem, and does the licence even apply in the EU?
01 · Context
Three media, one procurement problem.
What changed substantially from 2025 to 2026, and why the selection criteria differ from LLMs.
Image: multimodal models such as GPT-Image-2 and Google’s Nano Banana Pro have retired the old exclusion criterion “no legible text”. In the open-weights camp the licensing picture is finer than it looks: for Germany’s FLUX.2, only the smallest variant (klein 4B) is Apache 2.0, while klein 9B and dev carry a non-commercial licence. Successor FLUX 3 arrived in July 2026 but access is gated. Qwen-Image remains Apache 2.0 throughout. Tencent’s Hunyuan licences explicitly exclude the EU, UK and South Korea — the licence applies for Swiss companies, but not for German or Austrian ones.
Video: the biggest turning point is OpenAI’s retreat: per OpenAI’s deprecations page the Sora 2 API shuts down on 24.09.2026, with no successor named. The enterprise field goes to Google Veo 3.1 (SynthID, EU region Frankfurt), Runway and Adobe. Adobe deserves a closer look: Firefly has become a platform for 30+ third-party models, yet full IP indemnity still covers Adobe’s own models only. For on-prem, Alibaba Wan 2.2 and Lightricks LTX-2 are viable — again with a footnote: Wan 2.5 and 2.6 are closed, only 2.2 remains open.
Audio: the legal situation in music is moving, but more slowly than expected. Following the Warner–Suno settlement, licensed models with C2PA watermarking are gaining ground; Sony and UMG continue to litigate. A landmark ruling on fair use is not in sight for now — the schedule in the Suno case defers the decisive motions to 2027. More relevant in practice is a side effect of the Warner deal: generating a song grants a usage licence, not ownership. In voice, the market is shifting to realtime agents; GDPR-compliant on-prem transcription is technically well within reach using NVIDIA’s open models, and ElevenLabs has offered on-premise deployment since 2026.
02 ·
Media models: rights, provenance, on-prem
The dimensions on which media-AI procurement in Switzerland and the EU fails or stalls.
| Model | Type | Licence | Output rights | Provenance | Status | On-prem / hardware | Price (indicative) | DACH risk |
|---|---|---|---|---|---|---|---|---|
| Image – generation & design | ||||||||
| FLUX.2 familyBlack Forest Labs · Germany | klein 4B: · klein 9B + dev: FLUX Non-Commercial | klein & API free · dev needs paid licence | API: yes · : no | 11/25 · klein 01/26 · succeeded by FLUX 3 (07/26) | Yes · dev ~80 GB · klein from consumer GPU | API from $0.014/image (klein) to $0.07 (max) · weights free | EU vendor · check licence tiers | |
| FLUX 3Black Forest Labs · Germany · multimodal | API | Early access · open weights announced for later in 2026 | by agreement · gated early access | not documented | 23 Jul 2026 · video gated | weights for selected partners only | video $0.17/s (HD) · $0.29/s (Full HD) | EU vendor · limited availability |
| Stable Diffusion 3.5Stability AI · UK | Community License | free below revenue threshold | unclear | 3.5 · 10/2024 | Yes · consumer–high-end GPU | weights free · API credits | Getty: UK lost 11/25, US ongoing | |
| GPT-Image-2OpenAI · USA | API | free · () | 04/2026 | No | $0.006–0.21/image | |||
| Nano Banana ProGoogle (Gemini 3 Pro Image) · USA | API | free · on Vertex AI | (always on) | GA 05/2026 · 4K preview | No | $0.13/2K image · Imagen 4 from $0.02 | ||
| Firefly Image 5Adobe · USA | API | Adobe model: full · partner models limited only | () | 03/2026 | No · custom models possible | CC Pro $69.99/mo · API credits | comparatively low | |
| Midjourney V8.2Midjourney · USA | API | SaaS subscription | >$1M revenue: Pro/Mega plan required | No | V8.2 · 07/2026 | No | Basic $10 / Standard $30 / Pro $60 / Mega $120 /mo | Disney/Universal lawsuit · no |
| Recraft V4.1Recraft · USA | API | free on paid · brand/vector focus | No | V4.1 · 2026 | No | $0.04 · Pro $0.25/image | young vendor | |
| Muse ImageMeta (MSL) · USA | App | not documented | not documented | 07 Jul 2026 · Meta apps only | No | no API · in Meta AI, WhatsApp, Instagram | no API · rights and provenance unclear | |
| Qwen-Image-2512Alibaba · China | free · no thresholds | No () | 12/2025 | Yes · 20B, quantised from ~24 GB | weights free | · opaque training data | ||
| HunyuanImage 3.0Tencent · China · 80B-A13B | Hunyuan Community License | licence does not apply in EU, UK, South Korea — Switzerland unaffected | No | 09/2025 | (Yes · 80B , ≥3×80 GB) | weights free · not licensed for EU/UK/KR | legally excluded for DE/AT, permitted in CH | |
| Video – generation | ||||||||
| Veo 3.1Google DeepMind · USA | API | free · on Vertex AI | 01/2026 · 4K | No · EU region Frankfurt | $0.05–0.75/s | US vendor · EU region available | ||
| Sora 2OpenAI · USA | API | only until sunset | + Watermark | EOL 24.09.2026 | No | $0.10–0.70/s | service being discontinued | |
| Runway Gen-4.5Runway · USA | API | free on paid · enterprise contracts | 12/2025 | No | ~$0.25/s · plans from $12/mo | credit costs scale fast | ||
| Kling 3.0Kuaishou · China | API | free on paid · no | on free tier | 02/2026 · 4K | No | $0.08–0.17/s | CN data routing · check | |
| Seedance 2.0ByteDance · China | API | free on paid | global since 04/2026 (excl. USA) · 4K since 06/26 · 2.5 announced | No | ~$0.24/s (fal, 720p) | CN vendor · check /data routing | ||
| Firefly VideoAdobe · USA | API | Adobe model: · partner models limited only | () | 2026 | No | Premium $199.99/mo (first-party unlim., promo to 08/26) | comparatively low | |
| Luma Ray3 / Ray3.14Luma AI · USA | API | free from paid plans | no default | Ray3 · 4K/HDR · Ray3.14 since 26 Jan 2026 | No | Lite $9.99 / Plus $29.99 / Pro $90 / Unlimited $94.99 / Ultra $300 /mo | smaller vendor | |
| MiniMax Hailuo 2.3MiniMax · China | API | per API terms | not documented | 2.3 · 1080p/24fps | No | from ~$0.08/s (OpenRouter) · up to ~$0.50/s by tier | · check data transfer | |
| Wan 2.2Alibaba · China | (2.1/2.2 only) | free for 2.2 · newer versions only via Alibaba Cloud | No () | 2.2 · open (2025) · 2.5 + 2.6 API-only, 2.7 partly open | Yes · 14B, from 24 GB ( from 16 GB) | weights free | · own governance needed | |
| LTX-2 / LTX-2.3Lightricks · Israel | Community License (<$10M free) | free below $10M annual revenue · above that a licence agreement | optional | 01/2026 · 4K + audio · LTX-2.3 since 03/2026 | Yes · from RTX-4090 class | weights free · API via fal/Replicate | mind the threshold contractually | |
| HunyuanVideo 1.5Tencent · China | Hunyuan Community License | licence does not apply in EU, UK, South Korea — Switzerland unaffected | No | 11/2025 | (Yes · RTX 4090) | weights free · not licensed for EU/UK/KR | legally excluded for DE/AT, permitted in CH | |
| Audio – voice, music & transcription | ||||||||
| Eleven v3ElevenLabs · USA/UK | API | free from Starter · music (indie deals: Merlin/Kobalt) | music: · : classifier | v3 · 2026 | (AWS/GCP) · on-prem + on-device in early access | Starter $6 / Creator $22 / Pro $99 / Scale $299 / Business $990 /mo | compliance · | |
| GPT-Realtime-2OpenAI · USA | API | free (API terms) | No | 05/2026 | No · Azure EU possible | $32/$64 per 1M audio tokens | ||
| Cartesia Sonic 3.5Cartesia · USA | API | free from paid plans | No | ~90 ms latency | private deployments (enterprise) | ~$35/1M characters | young vendor | |
| Inworld Realtime TTS-2Inworld · USA | API | per API terms | not documented | -2 research preview · 1.5 generally available | on-prem on request | $15/1M chars (from $300/mo) · from $10/1M at volume | -2 still preview | |
| Azure AI SpeechMicrosoft · USA | free · gated by approval | available | HD-Update 03/2026 | containers: connected or disconnected () · EU regions | $16/1M (neural) · $22/1M (neural HD) · commitment from $7.50/1M | strong DACH compliance story | ||
| Fish Audio S2.1 ProFish Audio (OpenAudio) | Research License | commercial ONLY with licence agreement | No | S2 open since 03/2026 · S2.1 Pro current | Yes · 1 GPU | weights free (non-commercial) | licence ambiguity · cloning | |
| Kokoro 82MHexgrad / Community | free | No | #1 -Arena 01/26 | Yes · CPU/small GPU (~300 MB) | free · hosting costs in cents | no enterprise support | ||
| NVIDIA Canary / ParakeetNVIDIA · USA · STT | CC-BY-4.0 | transcripts freely usable | – | Canary-1b-v2 · Parakeet-v3 · 25 European languages | Yes · 1 GPU · NIM support | weights free · AI Enterprise plan optional | ideal on-prem transcription | |
| Suno v5.5Suno · USA · Music | API | usage licence from Pro — no ownership of the song; free tier has no download | -Watermark | licensed gen. 03/2026 | No | Pro $10 / Premier $30 /mo | UMG/Sony lawsuits ongoing | |
| Stable Audio 3.0Stability AI · UK · Music | Open weights (Small/Med) + API | free · fully licensed training data | unclear | 3.0 · 05/2026 | Yes · open weights (Small/Med) | usage-based · enterprise custom | comparatively licence-safe music option | |
| Workflow & orchestration – model-agnostic | ||||||||
| ComfyUIComfy Org · open source | Workflow tool | GPL-3.0 | depends on the loaded model | depends on the model | $30M at $500M valuation (04/2026, Craft Ventures) · cloud + API nodes | Yes · 1 GPU (model-dependent) | free (GPL) · Comfy Cloud optional | check GPL copyleft when embedding |
Image – generation & design
- Licence
- klein 4B: · klein 9B + dev: FLUX Non-Commercial
- On-prem / hardware
- Yes · dev ~80 GB · klein from consumer GPU
- Price (indicative)
- API from $0.014/image (klein) to $0.07 (max) · weights free
- DACH risk
- EU vendor · check licence tiers
- Licence
- Early access · open weights announced for later in 2026
- On-prem / hardware
- weights for selected partners only
- Price (indicative)
- video $0.17/s (HD) · $0.29/s (Full HD)
- DACH risk
- EU vendor · limited availability
- Licence
- Community License
- On-prem / hardware
- Yes · consumer–high-end GPU
- Price (indicative)
- weights free · API credits
- DACH risk
- Getty: UK lost 11/25, US ongoing
- Licence
- On-prem / hardware
- No
- Price (indicative)
- $0.006–0.21/image
- DACH risk
- Licence
- On-prem / hardware
- No
- Price (indicative)
- $0.13/2K image · Imagen 4 from $0.02
- DACH risk
- Licence
- On-prem / hardware
- No · custom models possible
- Price (indicative)
- CC Pro $69.99/mo · API credits
- DACH risk
- comparatively low
- Licence
- SaaS subscription
- On-prem / hardware
- No
- Price (indicative)
- Basic $10 / Standard $30 / Pro $60 / Mega $120 /mo
- DACH risk
- Disney/Universal lawsuit · no
- Licence
- On-prem / hardware
- No
- Price (indicative)
- $0.04 · Pro $0.25/image
- DACH risk
- young vendor
- Licence
- On-prem / hardware
- No
- Price (indicative)
- no API · in Meta AI, WhatsApp, Instagram
- DACH risk
- no API · rights and provenance unclear
- Licence
- On-prem / hardware
- Yes · 20B, quantised from ~24 GB
- Price (indicative)
- weights free
- DACH risk
- · opaque training data
- Licence
- Hunyuan Community License
- On-prem / hardware
- (Yes · 80B , ≥3×80 GB)
- Price (indicative)
- weights free · not licensed for EU/UK/KR
- DACH risk
- legally excluded for DE/AT, permitted in CH
Video – generation
- Licence
- On-prem / hardware
- No · EU region Frankfurt
- Price (indicative)
- $0.05–0.75/s
- DACH risk
- US vendor · EU region available
- Licence
- On-prem / hardware
- No
- Price (indicative)
- $0.10–0.70/s
- DACH risk
- service being discontinued
- Licence
- On-prem / hardware
- No
- Price (indicative)
- ~$0.25/s · plans from $12/mo
- DACH risk
- credit costs scale fast
- Licence
- On-prem / hardware
- No
- Price (indicative)
- $0.08–0.17/s
- DACH risk
- CN data routing · check
- Licence
- On-prem / hardware
- No
- Price (indicative)
- ~$0.24/s (fal, 720p)
- DACH risk
- CN vendor · check /data routing
- Licence
- On-prem / hardware
- No
- Price (indicative)
- Premium $199.99/mo (first-party unlim., promo to 08/26)
- DACH risk
- comparatively low
- Licence
- On-prem / hardware
- No
- Price (indicative)
- Lite $9.99 / Plus $29.99 / Pro $90 / Unlimited $94.99 / Ultra $300 /mo
- DACH risk
- smaller vendor
- Licence
- On-prem / hardware
- No
- Price (indicative)
- from ~$0.08/s (OpenRouter) · up to ~$0.50/s by tier
- DACH risk
- · check data transfer
- Licence
- (2.1/2.2 only)
- On-prem / hardware
- Yes · 14B, from 24 GB ( from 16 GB)
- Price (indicative)
- weights free
- DACH risk
- · own governance needed
- Licence
- Community License (<$10M free)
- On-prem / hardware
- Yes · from RTX-4090 class
- Price (indicative)
- weights free · API via fal/Replicate
- DACH risk
- mind the threshold contractually
- Licence
- Hunyuan Community License
- On-prem / hardware
- (Yes · RTX 4090)
- Price (indicative)
- weights free · not licensed for EU/UK/KR
- DACH risk
- legally excluded for DE/AT, permitted in CH
Audio – voice, music & transcription
- Licence
- On-prem / hardware
- (AWS/GCP) · on-prem + on-device in early access
- Price (indicative)
- Starter $6 / Creator $22 / Pro $99 / Scale $299 / Business $990 /mo
- DACH risk
- compliance ·
- Licence
- On-prem / hardware
- No · Azure EU possible
- Price (indicative)
- $32/$64 per 1M audio tokens
- DACH risk
- Licence
- On-prem / hardware
- private deployments (enterprise)
- Price (indicative)
- ~$35/1M characters
- DACH risk
- young vendor
- Licence
- On-prem / hardware
- on-prem on request
- Price (indicative)
- $15/1M chars (from $300/mo) · from $10/1M at volume
- DACH risk
- -2 still preview
- Licence
- On-prem / hardware
- containers: connected or disconnected () · EU regions
- Price (indicative)
- $16/1M (neural) · $22/1M (neural HD) · commitment from $7.50/1M
- DACH risk
- strong DACH compliance story
- Licence
- Research License
- On-prem / hardware
- Yes · 1 GPU
- Price (indicative)
- weights free (non-commercial)
- DACH risk
- licence ambiguity · cloning
- Licence
- On-prem / hardware
- Yes · CPU/small GPU (~300 MB)
- Price (indicative)
- free · hosting costs in cents
- DACH risk
- no enterprise support
- Licence
- CC-BY-4.0
- On-prem / hardware
- Yes · 1 GPU · NIM support
- Price (indicative)
- weights free · AI Enterprise plan optional
- DACH risk
- ideal on-prem transcription
- Licence
- On-prem / hardware
- No
- Price (indicative)
- Pro $10 / Premier $30 /mo
- DACH risk
- UMG/Sony lawsuits ongoing
- Licence
- Open weights (Small/Med) + API
- On-prem / hardware
- Yes · open weights (Small/Med)
- Price (indicative)
- usage-based · enterprise custom
- DACH risk
- comparatively licence-safe music option
Workflow & orchestration – model-agnostic
- Licence
- GPL-3.0
- On-prem / hardware
- Yes · 1 GPU (model-dependent)
- Price (indicative)
- free (GPL) · Comfy Cloud optional
- DACH risk
- check GPL copyleft when embedding
- ⚠ Adobe Firefly is a platform: Alongside its own models, Firefly offers 30+ partner models (including Google, OpenAI, Runway, Black Forest Labs, Kling). Full IP indemnity applies only to Adobe’s own models trained on Adobe Stock. Partner models fall under a limited supplemental coverage; beta, preview and early-access versions are excluded. If legal certainty is the reason for buying, the specific model in use must be checked (verified 7 Aug 2026).
- → Prices indicative: vendor list prices, before volume discounts.
- → Compiled editorially: as of 25 Jun 2026; points that could not be fully verified are rated conservatively.
- ⚠ Not legal advice: have licence and rights questions reviewed individually.
- → ComfyUI: listed as a model-agnostic workflow tool; output rights and provenance follow the model loaded into it.
Information in this market changes frequently and at short notice; provided without warranty and without claim to completeness. For binding terms, always consult the official vendor page.
Looking for coding agents and LLM tooling instead of media models? Open the AI coding tools comparison →03 · Context
Tooling around the models.
Beyond the models there is an ecosystem of interfaces and platforms, deliberately kept out of the matrix because they are not a model decision in themselves.
Workflow interfaces
ComfyUI is the de-facto standard for node-based media pipelines (hence the only tool in the matrix). Alternatives such as Invoke or SwarmUI can be easier for some teams, the model question stays the same.
Inference platforms
fal.ai, Replicate and the like host open models behind an API, fast start without your own GPUs, but data flows through the platform vendor. No answer for on-prem requirements, often the best one for prototypes.
Suite integrations
Adobe, Canva and increasingly DAM systems embed models directly into creative workflows. Convenient for teams, the rights and provenance questions from the matrix apply unchanged.
04 ·
Which model fits your initiative?
05 · Key terms explained
Key terms explained (23) +
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.
- air-gapped
- Operation with no internet connection at all, the highest isolation level for highly sensitive environments (e.g. critical infrastructure).
- VPC
- Virtual Private Cloud, a logically isolated area within a public cloud. More control than standard cloud, less than your own data centre.
- 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.