In 2026, AI video generation rapidly moved from 'technology for experimental play' to 'technology used in real work.' At the center of this shift is Seedance 2.0, a video-generation model developed by Seed, the AI research division of TikTok owner ByteDance.
Earlier AI video had commercially fatal weaknesses: characters' appearances changed partway through a clip, the physics looked wrong, and picture and sound fell out of sync. By resolving many of these problems together, Seedance 2.0 spread through video-production workplaces within only a short time of its release.
Combining Four Types of Material to 'Direct' Video
What changed technically? The biggest point is an 'integrated multimodal' design that can reference four kinds of material at once: text, images, audio, and video. A single generation can combine up to twelve assets in total—nine images, three videos, and three audio files.
This makes it possible to give highly detailed directions in words, such as: 'Use this image as the first frame, use this video as a reference for movement, and match the dialogue and sound to this audio.' The ability for a creator to actively 'direct' even the lighting, shadows, and camera movement, rather than waiting for a lucky result, is a major difference from earlier tools.
Creating Picture and Sound Together in a Single Flow
Another core feature is a system that creates video and audio together from the outset in a single pipeline. Previously, cumbersome postproduction was essential: first create a silent video, add dialogue in another tool, and then synchronize the mouth movement.
Because Seedance 2.0 generates picture and sound from the same foundation, lip movement closely matches the voice, and even ambient audio synchronized to the action emerges at once. It also includes an engine that reproduces physics such as gravity, the movement of clothing, and collisions, as well as the ability to recreate cinematic camerawork such as a dolly zoom or tracking shot with a single instruction.
Stable character consistency, in which faces are less likely to break down even amid intense movement, also encouraged professional adoption. Output supports high resolution equivalent to 2K and can generate clips of roughly fifteen seconds at a time.
Creators Use Repackaged Services, Rather Than the 'Core Model'
Seedance 2.0 itself, however, is a 'foundation model' developed by ByteDance and is not something ordinary creators use directly. In practice, access comes through a range of services that repackage the model for ease of use.
ByteDance itself provides it through CapCut and its dedicated Dreamina platform. Around those offerings, many unofficial wrapper services have appeared, using external APIs and providing credit-based pricing and accessible interfaces.
The quality of these services varies. Some receive poor ratings because their support and refund policies are unclear, so users need to evaluate each provider carefully.
'We Can Improve Every Day': Higgsfield as Orchestrator
One of the most strikingly successful services in this 'bridging' role is Higgsfield AI, which brings together multiple leading models. Founder Alex Mashrabov chose not to enter the competition to develop foundation models, which demands enormous computing resources. Instead, his strategy was to focus exclusively on being a 'workflow orchestrator.'
His view is as follows: because video generation requires far more computation than text generation, companies developing foundation models can see product-improvement cycles stretch to several months.
Workflow-focused platforms like his, by contrast, can improve their products almost daily on the basis of user responses—and this difference in iteration speed is the source of their competitiveness.
Higgsfield is in fact said to have reached an annualized revenue run rate of about $200 million roughly nine months after its founding, faster than companies such as Slack, Zoom, and Dropbox.
Mashrabov also reads the response of younger generations astutely. Today's viewers, he observes, already accept that a video itself was made with AI; rather than pixel-perfect realism, they seek the intuitive 'vibe' the video carries.
The irony is that delivering that vibe exactly as intended requires creators to exercise precise control over the model. This is precisely why detailed camera controls and directing features of the kind Seedance 2.0 provides have value. He has also offered the bullish forecast that, by the end of the 2020s, much of the video on social media will be AI-generated and a substantial share will be created through his company's systems.
From $100,000 per Minute to Hundreds of Dollars: Production Costs Collapse
These efficiencies have begun to shake the economics of video production itself. Broadcast-quality video was once estimated to cost around $100,000 per minute. Combining the latest models and orchestrators is said to have reduced that figure by orders of magnitude, to hundreds of dollars per minute.
There are also reports of highly complex shots that ordinarily would require enormous budgets being reproduced almost completely at minimal cost. AI video is therefore moving beyond prototypes and mockups toward a central role in production, including the direct creation of advertising assets and storyboards.
Competition with Sora 2 and Kling 3.0—and What Changed Afterward
The market structure also deserves attention. At the beginning of 2026, competition on the front line of AI video centered on OpenAI's Sora 2 and Kuaishou's Kling 3.0. While Sora 2 drew praise for the fidelity of its physics, Seedance 2.0 demonstrated strength in structured production workflows through its multimodal input, audio synchronization, and camera controls described above.
Developments moved quickly, however, and OpenAI changed direction in spring 2026 and ended availability of the Sora app. Companies also successively introduced low-cost, fast lightweight models, and the conventional assumption that speed, affordability, and high quality cannot coexist is beginning to break down.
Rights Holders Push Back, Prompting a Cautious Rollout Outside Europe and the United States
At the same time, this rapid progress caused intense friction with existing rights holders. SAG-AFTRA, the US actors' union representing about 160,000 media professionals, issued a statement strongly condemning Seedance 2.0, alleging that members' voices and likenesses were being used without either consent or compensation.
The conflict reached the political arena as well. Senators Marsha Blackburn and Peter Welch sent ByteDance CEO Liang Rubo a bipartisan letter calling for an immediate halt. In the first 24 hours after release, fake videos combining real actors and altered clips of popular television dramas had spread widely across social media, forming the background to the request.
In response to these headwinds, ByteDance adopted a cautious approach to rolling out Seedance 2.0. Initial availability was limited to countries including Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand, and Vietnam, with the European and US markets, where litigation risk was high, excluded from the outset.
At the same time, it introduced systems restricting the generation of real people's faces and copyrighted characters, as well as C2PA digital watermarking to indicate that outputs originated with AI. Some creators complained that this reduced their freedom, but it can be viewed as an unavoidable adjustment in the search for coexistence with the industry.
The Central Goal ByteDance Envisions: A 'World Model'
Another point that must not be overlooked is what ByteDance sees beyond this model. Wu Yonghui, who leads the Seed team, removed quarterly short-term targets from long-term fundamental research and created an environment in which researchers could settle in for sustained work.
The clear objective is a 'world model': technology that allows AI to simulate the workings of the real world itself. The company aims to release by the end of 2026 a world model comparable to Google's leading model.
Under Zhou Chang, who oversees multimodal systems and robotics, ByteDance also plans to apply the ability developed in video generation to connect 'vision, language, and action' beyond digital environments, using it as the intelligence of robots that move in the real world.
The simulations of gravity and collision refined in Seedance 2.0 are not solely for video. Ultimately, they also serve as a 'training ground' for moving robots safely in the real world.
The Difficult Questions That Remain Beyond Democratization
Seedance 2.0 presents a world in which video, audio, and the simulation of physics are integrated. It opened visual expression that had previously been within reach only of a few production companies with enormous budgets to individual creators and small organizations paying only a modest monthly fee.
At the same time, it highlights serious issues that technical measures alone cannot solve, including infringement of copyright and likeness rights and the livelihoods of human creators. The conversation about creating video with AI is already moving to the next stage: competition over world models that connect the digital and physical realms. How should we handle what is lost behind this convenience? It appears that all of us will have to seek the answer together.
