Seedance 2.5

Seedance 2.5 is ByteDance’s flagship video model. It generates video with synchronized sound from text, images, video clips, and audio, all in one pass. It takes mixed inputs and returns clips up to 30 seconds long in a single generation.

What’s new in 2.5

Seedance 2.5 builds on 2.0 with a focus on production work:

• Native 30-second generation. Create a full 30-second clip in one pass, with no stitching. Set duration from 4 to 30 seconds, or to -1 to let the model pick the length.
• Larger reference sets. Combine up to 30 images, 10 video clips, and 10 audio files in a single generation. Refer to them in your prompt as [Image1], [Video1], [Audio1], and so on.
• Better instruction following. It handles complex, multi-subject shots and prompts that lean on many references.
• Multilingual prompts. Write prompts in many languages.
• Adaptive aspect ratio. Set aspect_ratio to adaptive and the model picks the ratio that fits your inputs.

What you can create

Text to video

Describe a scene in natural language and get a video with matching audio. The model understands interactions between several subjects, camera movements, and emotional tone. For dialogue, put the speech in double quotes in your prompt, and the model generates matching lip movements and voice.

Image to video

Animate a still image by passing it as image, the first frame. You can also pass last_frame_image to control where the video ends. The model keeps the look and style of your image while adding natural motion. First-frame and last-frame runs need aspect_ratio set to adaptive.

Multimodal reference

Combine images, videos, and audio as references. For example, pass a reference video for motion style, reference images for a character’s appearance, and reference audio for rhythm, then describe how to combine them. This works well for outfit-change videos, product videos, and music-synced content.

Video editing

Pass a reference video and describe the changes, such as replacing an object, changing the background, or changing the style. The model keeps the original motion and camera work. Set aspect_ratio to adaptive and duration to -1 for editing, because the output follows the input video.

Video extension

Pass a reference video and describe what should happen next. The model continues the scene with the same characters, setting, and style. Set aspect_ratio to adaptive for extension.

Key features

Native audio generation

The model generates audio and video together, so dialogue, sound effects, and background music line up with the picture from the start. Set generate_audio to false if you want silent video.

Character consistency

With reference images, the model keeps faces, clothing, and style the same across the video. You can use this to make multi-shot stories with consistent characters.

Precise prompt following

The model handles complex prompts with several subjects, specific actions, and detailed camera movements. It understands spatial relationships and actions that happen in sequence.

Tips for good results

• Be specific in your prompts. Describe camera movements, lighting, mood, and actions.
• For dialogue, put the spoken words in double quotes: The man stopped and said: "Remember this moment."
• When you use reference inputs, label them in your prompt: “The character from [Image1] performs the dance from [Video1].”
• For video editing, say what to change and what to keep: “Replace the perfume in [Video1] with the face cream from [Image1], keeping all original motion.”
• Start with 5-second clips while you experiment, then make them longer once you like the style.

Learn more

For technical details on the Seedance 2 family, see ByteDance’s Seedance 2.0 page.

You can try this model in the Replicate playground.