NOTE FOR EU COMPANIES: Seedance comes from ByteDance (China). Via the consumer apps Dreamina and CapCut there is still no documented GDPR-compliant EU data processing. Via the API of ByteDance’s international subsidiary BytePlus (ModelArk), however, a dedicated EU region in Dublin (Ireland) now exists – though per BytePlus documentation without an absolute routing guarantee. On top of this, the model sits at the centre of significant copyright and deepfake controversies. For professional use in the DACH region, we still recommend a careful data-protection and compliance review.
innFactory AI Consulting from Rosenheim assesses new video-AI models for use in the DACH region – with a focus on data protection, compliance and practical suitability.
What is Seedance 2.0?
Seedance 2.0 is ByteDance’s (team “ByteDance Seed”) native multimodal audio-video generation model. Unlike pure text-to-video systems, it processes four input modalities – text, image, audio and video – covering text-to-video, image-to-video, reference-to-video, video editing and video extension. The distinctive part: audio (speech, sound effects, music) is generated natively and in sync within the same model. This overview reflects the state as of October 2026.
The model was first released in China in early February 2026; a technical report from the 200-plus-person team appeared on arXiv (2604.14148) on 15 April 2026. The global rollout to more than 100 countries including Europe – but excluding the US (copyright dispute) – followed from late March/April 2026, via Dreamina and the CapCut integration, among others.
Technical details
- Resolution: natively 480p and 720p. A later extension to 4K was reported for the Volcano Engine conference in June 2026, but so far is only documented via secondary sources and was not a launch feature.
- Clip length: 4 to 15 seconds of direct audio-video generation.
- Multimodal references: up to 3 video clips, 9 images and 3 audio clips can be combined in a single run.
- Audio: binaural, synchronised audio tracks with strong lip-sync (background, ambient, narration).
- Multi-shot: native storytelling across multiple shots with camera and cut direction (“cinematographic reasoning”).
- Architecture: a “unified, highly efficient architecture for multimodal audio-video joint generation”. The exact architecture type is not explicitly named in the report.
- Variants: Seedance 2.0 Fast (low-latency) and Seedance 2.0 Mini (cheaper draft stage, June 2026).
What’s new versus Seedance 1.5?
Seedance 1.5 Pro already introduced synchronised audio-video generation within the Seedance line. Version 2.0, per the report, takes the step to a unified architecture for jointly generating image and sound. The key advances:
- markedly better motion and physics plausibility and temporal coherence
- stronger instruction-following, including for long scripts
- better character and subject consistency across multiple shots
- new editing capabilities (targeted changes without full regeneration)
In ByteDance’s own benchmark, 2.0 improves on the predecessor 1.5 by an average of 0.86 points (scale 1–5), with the biggest jump in motion quality.
Benchmarks – to be read with caution
According to ByteDance’s own technical report, Seedance 2.0 takes first place in text-to-video and image-to-video in an arena snapshot (April 2026), ahead of Google Veo 3.1 and OpenAI Sora 2 Pro, and leads all six dimensions in the in-house “SeedVideoBench 2.0” benchmark.
These figures should, however, be put in context: the arena entry cited in the paper was flagged as “preliminary” with a small sample (around 2,700 votes), and “SeedVideoBench 2.0” is a vendor-owned benchmark. A top ranking is, however, also confirmed by the independent portal Artificial Analysis, which at times placed Seedance 2.0 first in its text-to-video ranking. Such rankings fluctuate, though – in April 2026 Seedance 2.0 was briefly displaced by another model – so a lasting “No. 1” is not a reliable claim.
Availability and cost
Seedance 2.0 is available via ByteDance’s platforms Doubao and Jimeng (China), the Volcano Engine/Ark API, and internationally via Dreamina and CapCut. Billing through Volcano Engine is token-based; a rule of thumb of around 1 RMB (~$0.14–0.15) per second is cited (secondary source). The target audience is content creators, advertising/marketing, film/TV effects, game animation and social media.
Data protection, copyright and risks for EU customers
For companies in the DACH region, provenance is decisive: ByteDance is a Chinese corporation (also the operator of TikTok). For the consumer access points Dreamina and CapCut we still have no robust assurances from primary sources of GDPR-compliant EU data processing.
Via the API of ByteDance’s international subsidiary BytePlus, ModelArk provides an enterprise access point that, per BytePlus documentation, offers a dedicated EU region in Dublin, Ireland (region ID eu-west-1, base URL ark.eu-west.bytepluses.com) alongside an Asia-Pacific region in Johor (ap-southeast-1). There, per BytePlus, API keys and model activation are managed with regional isolation, and an inference endpoint created in one region can only be reached via that region’s base URL. Important caveat: BytePlus itself documents that requests are processed primarily in the chosen region, but some may be routed to other regions based on resource scheduling – this is not an absolute EU data-residency guarantee in the sense of strict compliance requirements. BytePlus is also formally the internationally (Singapore-)based ByteDance subsidiary, but part of the same corporate group as the Chinese ByteDance parent; DPA, the subprocessor list and intra-group data flows should be clarified directly with BytePlus before production use involving personal data.
On top of this comes a significant copyright and deepfake controversy: in early 2026 ByteDance faced cease-and-desist demands from Disney and Paramount Skydance, a protest from the actors’ union SAG-AFTRA, and a letter from US senators – triggered by viral, realistic deepfakes of celebrities. ByteDance responded with safeguards: C2PA watermarking for provenance, live verification when uploading real people, proactive IP monitoring and external red-teaming.
Successor: Seedance 2.5
In June 2026 (Volcano Engine FORCE conference on 23 June) ByteDance announced the successor Seedance 2.5: native 30-second one-shot videos, up to 50 multimodal references and “long-form shot consistency”. The public launch took place in early July 2026, initially as a global enterprise beta.
Seedance 2.5 is usable via BytePlus ModelArk with published API documentation (model ID dreamina-seedance-2-5-260628), no longer just a closed enterprise beta. Per BytePlus pricing documentation, a 28% introductory discount on 1080p output applied from 14 August to 17 September 2026 (from around USD 0.41/second); this discount has expired. We could not confirm the widely reported 4K capability in the BytePlus documentation we reviewed – it references 480p, 720p and 1080p output.
Our recommendation
Per ByteDance, Seedance combines native audio generation and multi-shot storytelling in one model; Seedance 2.5 adds clips of up to 30 seconds. For productive use in the DACH region, data-protection and compliance questions remain in our view: the consumer access points Dreamina/CapCut still have no documented EU data residency, and even BytePlus ModelArk’s new EU region in Dublin offers no absolute routing guarantee per the vendor. On top of that come the IP risks around the deepfake controversy. Anyone who still wants to work with Seedance should prefer the BytePlus ModelArk API access with an explicit EU region choice (eu-west-1) over direct consumer access via Dreamina/CapCut, and review the DPA and subprocessors beforehand. For data-protection-sensitive companies, we recommend options where processing in the EU is documented or under your own control:
- Lightricks LTX: open weights for video with sound, self-hosting in the EU
- Black Forest Labs FLUX: FLUX 3 Video (preview) via the BFL API with its EU endpoint
Per Google’s model pages, Google Veo is available on the Agent Platform only in us-central1 (USA), and OpenAI Sora was shut down in September 2026.
For an assessment of which video-AI solution fits your data-protection and compliance requirements, contact innFactory AI Consulting.