ByteDance announced Seedance 2.5, its new AI video generation model, on June 23 at its Volcano Engine FORCE conference in Beijing. The headline feature is the ability to generate a 30-second clip in a single pass, without stitching shorter segments together. The model also accepts up to 50 reference materials and supports localized editing that redraws only part of a frame. It is currently in enterprise beta, with a public launch planned for early July.

Why a Seamless 30 Seconds Matters for Video Production

Until now, major AI video generators have been limited to roughly 8 to 15 seconds per single generation. Google's Veo 3.1, Runway's Gen-4.5, and OpenAI's Sora 2 (now discontinued) all sat within that range, and producing longer footage required stitching short clips together in sequence. At those seams, however, a character's face could shift subtly, lighting and color could fail to match, and the result often betrayed its AI origins.

Many TV commercials, social media ads, explainer videos, and short documentaries fall within the 15-to-60-second range. ByteDance says Seedance 2.5 can produce that span as a single seamless clip. Rather than generating video and audio separately and syncing them afterward, the company processes both within the same internal space, and says its proprietary attention mechanism for temporal consistency holds character, lighting, and motion style steady across the full 30 seconds.

Up to 50 Reference Inputs, Plus Editing Just One Part

Seedance 2.5 offers three generation modes: text-to-video, which builds a clip from a written prompt; image-to-video, which animates a still image; and a motion-reference mode that borrows camera work and movement style from existing footage.

The reference handling draws the most attention. Seedance 2.5 can take in up to 50 images, video clips, and audio files at once. That is a sharp increase from roughly 12 in the previous Seedance 2.0, and the gap is notable given that Google's Veo 3.1, the closest competitor for reference control, accepts only 3. Volcano Engine president Tan Dai described it as a feature aimed at brand work and episodic production, where a character's face, a product's look, an art style, and a musical mood must stay consistent across an entire clip.

The model also supports localized editing that redraws only a specific element in the frame. A character's clothing, a background, or a product in a scene can be swapped without regenerating the whole clip. Previously, fixing a small error meant starting over and hoping the rest of the output survived. Reducing that rework pays off in practice.

Seedance 2.0 Goes 4K and Leads an Independent Ranking

At the same event, ByteDance also revealed that the current Seedance 2.0 now supports native 4K output with 10-bit color depth, an upgrade for teams already building on 2.0.

As evidence of its strength, Seedance 2.0 ranks first in the text-to-video-with-audio category on the Artificial Analysis Video Arena, an independent leaderboard based on blind human preference, with an Elo score of 1,218, ahead of Kling 3.0 and Veo 3.1. That score reflects the shipping 2.0; no score exists yet for 2.5. ByteDance also said its enterprise Seedance business has reached 2 billion USD (about 320 billion yen) in annual recurring revenue, a sign of commercial traction ahead of the 2.5 launch. Distribution runs through the company's Volcano Engine cloud, along with consumer-facing channels such as the video editing app CapCut, which has more than 400 million monthly users, plus Dreamina and Doubao.

*1 USD = 162 JPY (as of June 25, 2026)

The Open Questions: Independent Testing and Copyright

Every capability described here remains ByteDance's own claim as of the June 23 announcement. The 30-second single-pass generation, the 50-reference handling, and the stability of localized editing have not been verified by third parties. That first outside scrutiny will come after the expected early-July public launch. Pricing has not been disclosed.

The copyright history also bears watching. After Seedance 2.0 launched in February 2026, the Motion Picture Association sent ByteDance its first-ever cease-and-desist letter to a major generative AI company. Disney, Warner Bros. Discovery, Paramount, Netflix, Sony Pictures, and Universal followed with their own letters, and ByteDance paused the global rollout in March. The company has since added filters that block the generation of real people's faces and copyrighted characters, and those safeguards are expected to carry over to 2.5. At FORCE, ByteDance also previewed a copyright licensing and monetization platform with filmmaker Stephen Chow as an initial partner. Because ByteDance is headquartered in Beijing, teams handling sensitive material should also review the data-handling rules of their own jurisdictions before use.

Summary

Seedance 2.5 arrives claiming it can generate a 30-second clip in one pass without stitching, paired with practical features like up to 50 reference inputs and localized editing. With Seedance 2.0 already topping an independent ranking, the model could pull ahead in the difficult area of long-form video generation, if its specifications hold up. For now, though, these are the company's own announcements: independent verification is still to come, and the copyright questions remain only half-settled. How its capabilities hold up at the July public launch will be the first thing worth watching.