On June 23, 2026, China’s ByteDance announced its flagship video-generation model, Seedance 2.5, at FORCE, a conference held in Beijing by its Volcano Engine cloud division.
The model is currently in beta for global enterprises, with general availability planned for early July. Seedance 2.0 arrived in February 2026, so this update follows only a few months later, but its substance goes far beyond a minor refinement.
Up to 30 seconds in a single generation, native 4K, and as many as 50 reference assets: these three advances push AI video beyond the “animated GIF” stage and closer to footage that can be used in production.
Up to 30 Seconds in One Continuous Take—Generation without Seams
Duration has been the greatest obstacle for AI video. To avoid breakdowns, many tools cap generations at roughly 5 to 15 seconds, forcing creators to assemble longer scenes from short clips.
Those joins have created major friction in professional production. Seedance 2.5 can directly output up to 30 seconds from one prompt and one generation.
Moving from 15 seconds to 30 is more than simply doubling the length. Fifteen seconds allows only a brief movement or highlight, while 30 seconds can contain an entire cinematic shot: establishing the situation, developing the action, following it with the camera, and leaving an afterglow.
A scene can reach its conclusion without being cut off midway. That is the key change turning AI video into a practical shot generator.
Native 4K with More Precise Physical Detail
The model also supports native 4K, and the earlier Seedance 2.0 has received a 4K update as well. This means more than simply adding pixels.
Fine physical simulations have become much more precise, from moving animal fur and macro time-lapses of flowers opening to fluids such as molten glass and smoke, and even the airflow around a bird’s beating wings.
Edges remain sharp and surface detail is retained. The result is designed to withstand demanding production requirements such as projection on large LED walls for filmmaking and high-resolution compositing.
Up to 50 Reference Files: From a Lottery to Deliberate Design
For professional creators, the largest change is the reference-asset limit. Seedance 2.0 accepted up to 12 reference files; version 2.5 raises that number to 50.
It can simultaneously ingest every type of material: images, video clips, audio, 3D white models that specify composition and lighting, and style references that define a particular visual treatment. This changes the nature of generation itself.
Generation based only on text prompts was a probabilistic process, much like a lottery in which the result was left to the model’s interpretation. Supplying 50 references and tightly guiding the model moves it closer to deterministic rendering that faithfully follows the provided design materials.
It can now maintain the appearance of the same person or product across multiple generations and support fine control such as replacing only one portion without disrupting the overall style.
Doubao 2.1 Pro Announced Alongside It—ByteDance Builds a Full Stack
The announcement was not limited to Seedance 2.5. ByteDance is assembling a broad front in the cloud market, and the same conference also introduced Doubao 2.1 Pro, the latest version of its large language model.
ByteDance says it has made major advances in coding, agent capabilities, and visual-language understanding, and claims that the model surpasses Anthropic’s Claude Opus 4.6 on several benchmarks.
These are ByteDance’s own claims and have not yet been independently verified. Pricing is highly aggressive: 6 yuan per million input tokens and 30 yuan per million output tokens, with cached input falling to 1.2 yuan. Volcano Engine estimates total cost of ownership at roughly 80% below Claude Opus 4.6.
Calls to the Doubao family now exceed 180 trillion tokens per day, more than ten times the volume of a year earlier. According to research firm IDC, Volcano Engine holds about 49.5% of China’s public-cloud MaaS, or model-as-a-service, market.
Seedream 5.0 Pro for image generation and Seed-Audio 1.0 for audio generation were previewed at the same time. ByteDance is evolving into a full-stack AI provider offering high performance and low prices across text, images, audio, and video.
Pricing and Access Vary by Use: Jimeng, Dreamina, and the API
Seedance is notable not only for performance but also for how it is distributed. For creators in China, the official Jimeng app is the main route. Deep integration with Douyin and the Chinese version of CapCut, with basic monthly plans starting at roughly a little over ¥1,000, makes it unusually affordable access to a leading video model.
Use requires a Chinese phone number and domestic payment methods such as WeChat Pay or Alipay, creating a practical barrier for overseas users. Dreamina, an official English-language platform that accepts international credit cards, is available to users without a Chinese number, but costs more than Jimeng and offers tiered higher-end plans.
Volcano Engine also offers an API for enterprises that want to integrate high-volume generation into their own systems. Standard-model generation costs roughly one yuan per second. Seedance 2.0 Mini, a lightweight version optimized for speed and cost efficiency, is available for substantially less than the standard model.
What the 95-Minute AI Film “Hell Grind” Reveals about Production
The clearest account of what this model family brings to production is “Hell Grind,” a feature-length AI film screened in Cannes in May 2026. Created by San Francisco startup Higgsfield AI using its own tools and Seedance 2.0, the 95-minute science-fiction heist action film moves between settings ranging from a Tibetan temple to feudal Japan.
It was not part of the Cannes Film Festival’s official selection. It was a third-party screening held around the industry market taking place at the same time, and the festival itself stated that it was unofficial.
The production data is striking. A team of only 15 completed the 95-minute feature in 14 days. A comparable traditional Hollywood production would require hundreds of people, several years, and tens of millions of dollars. The total budget was just under $500,000, yet around 80%, or $400,000, went to generative-AI compute rather than actors or locations. To create the first 25 minutes, the team generated 16,181 clips and ultimately used only 253 shots.
That is a yield of 64 to 1. Individual generations may be cheap, but finding a perfect shot that satisfies the director’s intent—gravity behaving correctly, props not floating, and performances remaining continuous—requires enormous trial and error, producing a large compute bill.
Thirty-Second Generation and 50 References Address Real Production Pain
Seen in this context, Seedance 2.5’s native 30-second generation and support for up to 50 reference files are responses to pressing production problems. Stronger control through many references could greatly improve that 64-to-1 ratio and reach the intended image in fewer takes.
The change also challenges the shape of video-production work itself. Execution—operating cameras, building lighting setups, and managing crews—will increasingly become commoditized, putting substantial pressure on mid-level production staff.
Value, meanwhile, will shift from how to make something toward what should be made, and toward the curation and direction required to select the best result from countless candidates.
Applications Expand from Film to Architecture and Industry
The impact extends beyond film. In architecture and spatial design, creators can feed a few stills exported from drawings as starting and ending frames, instantly producing a slow journey through an interior or a time-lapse in which the light shifts from day to night.
It can also create a smooth transformation from an empty room to a furnished one. Complex camera moves orbiting a subject can still cause spatial breakdowns, however, so the technology is not universal.
Industrial applications are also advancing, including virtual presenters for live commerce and synthetic training data for autonomous driving.
An Offensive Move: A Platform for Commercializing AI Copyright
More surprising to the industry than the technical specifications was ByteDance’s simultaneous unveiling of an “AI Copyright Commercialization Platform.” The generative-AI industry’s largest weakness is the risk of copyright infringement and the backlash against training on and generating creators’ work without permission.
ByteDance answered with an offensive design rather than a defensive system that blocks results after generation: secure formal licenses in advance, then build lawful derivative creation and revenue sharing into the platform.
A leading initial partner is Hong Kong filmmaker and actor Stephen Chow. Using authorized templates in ByteDance’s tools, users can create derivative works based on scenes from Chow’s classic films without fear of infringement.
Works using the relevant templates reportedly reached tens of thousands on the day of the announcement, demonstrating the speed of adoption. At the core is a model in which revenue generated when a work is viewed is shared among ByteDance, the creator, and the rights holder.
If it succeeds, rights holders may come to see AI as a system that continually produces new licensing income from older intellectual property, rather than as a threat that takes away their rights. It is a move that establishes a legal and commercial moat ahead of competitors focused only on performance.
The Main Contest Shifts from Image Quality to Ecosystems
Seedance 2.5 is therefore more than a feature update: it marks AI video’s serious move into professional production. Native 30-second generation, detailed 4K, and support for 50 reference files move video creation away from probabilistic, prompt-dependent generation and toward deterministic rendering grounded in detailed references.
On-the-fly synchronization of audio and video is another Seedance strength, but competitors such as Google Veo 3.1 and Kuaishou Kling 3.0 also provide native audio, so competition remains close in this area.
The main arena for video generation is shifting from a purely technical contest over which model makes the most beautiful images to a platform contest over which can build the safest ecosystem that benefits both rights holders and creators.
Once the service formally launches in July, creators will increasingly need the ability to build references that guide AI output precisely and the judgment to weave a story from a sea of generations, more than skill in operating cameras and lights.
