Seedance 2.5 by ByteDance: 30 Seconds in One Pass

Seedance 2.5 release date and changes since Seedance 2.0

ByteDance's Seed team launched Seedance 2.5 on July 31, 2026, and the post highlights three changes: 30 seconds per pass with multi-round extension, multimodal reference and timestamp-level editing. The post says users now expect a video model to finish a piece of work, not only to generate a clip. The model builds on the audio-video joint-generation architecture of Seedance 2.0; the post says nothing more about what 2.0 could do.

30 seconds per pass, then multi-round extension to minutes

A single pass means one generation, with no clips spliced together afterward. Seedance 2.5 lets one pass run 30 seconds with audio, up from 15.

The post says the model can fit several connected shots into those 30 seconds, so a story moves through setup, development, turning points and resolution instead of stretching one moment. Its example is a singer who deals with staff in the dressing room, walks the backstage corridor, meets the dancers and steps onto the stage, all in one take.

Multi-round extension goes past 30 seconds. A user appends later shots to an existing video, and the model keeps the main characters, environments and narrative pacing consistent. The post says this yields videos lasting several minutes and reduces the effort of splitting clips, splicing footage and fixing transitions.

The post also lists quality fixes. Camera moves transition more smoothly, the main subject stays stable across cuts, and audio stays in sync with the picture. ByteDance tuned object textures, skin and eye features, lighting and color saturation, and says the model now adds fewer subtitles and less background music that nobody asked for.

Seedance 2.5 reference limits: 30 images, 10 clips, 10 audio

Seedance 2.5 takes up to 30 images, 10 video clips and 10 audio clips as reference in one pass. Reference means handing the model real material and telling it what to use from each piece. The post says the model reads composition, scenes, styles, characters and props across all of it, and keeps each character's look, voice and traits stable in shots with several characters.

Clay render reference sets composition and camera angles

A clay render, a set of textureless 3D models, lets users fix a scene's composition and camera angles before the model generates anything. A creator who wants a car assembly sequence to follow set camera work builds the parts and their positions as a clay render first, then supplies an image for materials, lighting and color. One of the post's example prompts describes exactly this.

With a clay render, users fix spatial structure, character poses, motion paths and camera angles. The model then builds the video from that structure, so composition and blocking (where each figure stands and moves) match the plan. The post adds that the model uses the clay render's spatial information to make the lighting follow physics: light source direction, color temperature, intensity and shadow projection.

Timestamp editing and green screen changes after generation

Timestamp-level control lets a prompt or an edit target a specific stretch of the video, both during generation and after it. During generation, a prompt can control the narrative, camera perspective, movement and overall rhythm for a given time frame. After generation, users can change characters, actions or plot elements in specific clips while keeping continuity and realism before and after the edit.

Green screen editing swaps the background and tells a different story while the main subject stays intact. The post says the model renders how the subject responds to the physical rules of the new environment. Its examples are the direction clothes flutter, how the hair behaves, the rhythm of a walk and how light hits the subject.

The post also names camera perspective editing and reference-based editing, aimed at professional work in film and advertising. The example prompts mark time ranges in seconds. One green screen prompt splits a 15-second plan into 0-4s, 4-10s and 10-15s, and another edits an existing video and changes only the camera movement.

Education, manufacturing and driving uses ByteDance cites

Beyond creative video, the post lists four fields where ByteDance says Seedance 2.5 is moving into real use. These are ByteDance's own claims; the post names no outside customer and gives no result figure.

Can you use Seedance 2.5 on Jimeng AI, Doubao or an API?

Yes on Jimeng AI and Doubao Pro, where the July 31 post says Seedance 2.5 is rolling out; the API is described as coming soon via BytePlus ModelArk, with no date. To try it, open Video Generation on Jimeng Web or in Doubao Pro and select Seedance 2.5. See also MiniMax H3 Open Weights: 768p Only, US/EU Excluded.

• Jimeng Web, then Video Generation, then Seedance 2.5
• Doubao Pro, then Video Generation, then Seedance 2.5

Pricing, free access and 4K: what the blog leaves out

The post states no price, no free tier, no resolution figure including 4K, no API date, no list of supported countries and no comparison with other models. Anyone who needs one of these has to look beyond the post.

Weaknesses ByteDance names: complex motion, many subjects

ByteDance names two weaknesses: the physical plausibility of complex motions and the stability of scenes where several subjects interact. The post says there is still room for improvement in both. Its stated direction is more coherent storytelling, a more intuitive generation and editing experience, and a deeper grasp of real-world physics.

ByteDance frames Seedance 2.5 as a move from generating single clips to handling the whole creative workflow. The three changes it highlights are 30 seconds per pass, larger references and timestamp editing.