Seedance 2.5 AI video generator

Seedance 2.5 is a video model from ByteDance’s Seed team, released on July 31, 2026, as the successor to Seedance 2.0. It generates picture and sound together in clips of 4 to 30 seconds, from a text prompt and reference images, video or audio, and it can cut between linked shots inside one clip. On DreamBench it renders at 480p, 720p or 1080p, with sound on or off.

There are 2 ways. On the Video page, pick Seedance 2.5, write your prompt and choose a length from 4 to 30 seconds: your words go to the model as written and come back as one clip. In Film Studio, describe the video you want and add reference images if you have them: DreamBench plans the scenes, has Seedance 2.5 make each one and joins them, and you can redo any scene.

Seedance 2.5 is priced in credits for each second of video: least at 480p, more at 720p and most at 1080p, and the same with sound on or off. A clip on the Video page adds a small base amount; a film adds a base amount for planning it. Credits come with a paid subscription, Seedance 2.5 is on all of them, no ByteDance account or API key is needed, and you see the exact cost before you generate.

Seedance 2.5 doubles the longest clip from 15 to 30 seconds and takes up to 30 reference images, 10 video clips and 10 audio clips, where 2.0 takes 9, 3 and 3. It follows timings written in whole seconds, where 2.0 follows shot numbers, and ByteDance reworked its skin, eyes, texture and light, so the same prompt looks different on each. Both generate sound and can cut between shots inside a clip.

One Seedance 2.5 clip runs 4 to 30 seconds. On DreamBench’s Video page you get exactly that: one clip, at the length you pick. In Film Studio a video can run longer, up to 30 seconds on Starter, 2 minutes on Plus and 5 minutes on Ultra and Studio, because DreamBench cuts it into scenes of 3 to 12 seconds and joins them. ByteDance’s own tools can also extend a clip.

Seedance 2.5 renders at 480p, 720p or 1080p on DreamBench. All 3 are on every subscription, and 1080p is the highest the model renders. 480p costs the least per second, so it suits a first try of an idea or a timing; 720p suits a social clip watched on a phone; choose 1080p for a final cut or a bigger screen. Pick the resolution in the composer before you generate.

On DreamBench there are 5: square 1:1, widescreen 16:9, vertical 9:16, classic 4:3 and cinematic 21:9. Pick one in the composer, or name it at the start of your prompt and DreamBench sets it for you. ByteDance’s API also lists a taller portrait frame and an adaptive setting that follows the prompt or the input picture; DreamBench doesn’t offer those for Seedance 2.5.

Yes. With sound on, Seedance 2.5 makes dialogue, effects, ambience and music with the picture, timed to what happens on screen; switch it off for picture only, at the same price. BytePlus lists 11 languages for prompts and speech: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean. Put each spoken line in quotation marks and name the language it is spoken in.

Yes. One Seedance 2.5 clip can cut between linked shots, and ByteDance pitches it at a short story with a set-up, a turn and an ending, the main subject keeping its look across the cuts. Number the shots and give each its seconds, such as “Shot 2 (4–8 s)”. On DreamBench’s Video page your shots reach the model as you wrote them; in a film, a scene’s clip can cut too.

On DreamBench, up to 8 per video, each tagged as a character, creature, location or prop and named in your prompt with @ and its label. DreamBench builds a sheet from each and sends it with every scene it appears in. ByteDance’s API takes up to 30 images, 10 video clips and 10 audio clips in one generation. The Video page takes words only, so references belong in Film Studio.

No. ByteDance doesn’t let Seedance 2.5 take a reference that shows a real person’s face, to protect people’s likeness, and DreamBench suggests another model when one is refused. ByteDance accepts faces from its own models, from preset digital characters and from real people who verified and authorized their likeness. So describe the person instead: if Seedance flags a character DreamBench designed from your words, DreamBench registers it with ByteDance’s asset library and tries again.

Yes. The model can start a clip from an image as its first frame, or work from reference images of a character, product or place. On DreamBench, add your images in Film Studio and tag them: DreamBench builds a sheet from each and sends it with every scene, so your product, place or character appears where your prompt puts it. The Video page takes words only, and no reference may show a real face.

The model does: ByteDance’s API takes 1 image as the first frame, or 2 as the first and last frames, and the video takes the first image’s shape. DreamBench’s composer has no place to add your own first or last frame. Add the picture as a reference image instead and DreamBench builds the scenes from it, and describe in your prompt how the clip should end.

On DreamBench, videos made on Starter carry a small DreamBench watermark, and Plus, Ultra and Studio remove it. ByteDance’s API can add a mark in the lower-right corner saying a video is AI-generated, an option that stays off unless asked for, and DreamBench doesn’t ask for it. The watermark depends only on your DreamBench subscription, not on the resolution or the length.

Yes. DreamBench’s Terms of Use allow personal and commercial use of what you make, including ads, product videos and client work. You need the rights to anything you upload, such as product photos or a logo, and generated videos may not be exclusive to you or protected by copyright. Starter videos carry a DreamBench watermark, which Plus and above remove.

ByteDance suggests a one-line summary first, then the action shot by shot or second by second, then what stays the same throughout. Open with the frame and the length, number the shots with their seconds, give each a size and a camera move, describe the light and the look, and end with the sound, spoken lines in quotation marks. Ask for “no subtitles” or “no music” if you want none.

Usually too much happens in too little time: ByteDance warns that a crowded stretch can come out with extra cuts or missing beats. Give each shot one main action and a few seconds, keep the motion simple, and avoid several people interacting at once, which ByteDance says is not yet fully stable. Lettering can come out misspelt. In a film, redo the scene that missed and keep the rest.

Inside one clip, describe the subject once and add a line asking for the same subject, place and light in every shot. Across a film, give DreamBench a reference image of each character, tagged as a character, and name them with @ in your prompt; DreamBench builds a sheet from each and sends it with every scene they appear in. Keep each character’s description the same throughout.

ByteDance’s API has one: it makes a 480p preview, then turns that same preview into a final video. DreamBench doesn’t use it. Instead, 480p is a resolution you can choose for any Seedance 2.5 video, at the lowest price per second, to test an idea; then generate it again at 720p or 1080p. That is a new take, so the idea carries over but the details can differ.

They do different jobs. Seedance 2.5 runs to 30 seconds a clip and can cut between shots inside it; Kling 3.0 runs to 15 seconds in one continuous take, with lip-synced dialogue in 5 languages; Veo 3.1 runs to 8 seconds, always with sound, in widescreen or vertical. Seedance 2.5 and Veo 3.1 are on the Video page, and Kling 3.0 runs in Film Studio.

Available on every plan. Video length depends on the plan: up to 30 seconds on Starter, 2 minutes on Plus and 5 minutes on Ultra and Studio.

Yes. Every film on this page is a published DreamBench template made with Seedance 2.5, shown with the prompt that made it. No film or still from another model appears on this page.