Seedance 2.5: ByteDance's 30-Second One-Take Video Model

Key facts at a glance

Developer

ByteDance Seed team

Release date

July 31, 2026

Clip length

4–30 s per request, extendable

References

Up to 30 images, 10 videos, 10 audio

Resolution

480p/720p 8-bit, 1080p 10-bit, no 4K

Model ID

dreamina-seedance-2-5-260628

Open source?

No, closed weights

Uncensored?

No, moderated service

What is Seedance?

It is the latest model in ByteDance Seed's Seedance video generation series, released on July 31, 2026. It builds on the unified audio-video architecture of Seedance 2.0 and focuses on three things: longer storytelling, richer references and finer editing. ByteDance offers it in its own creative apps such as Dreamina and to developers through BytePlus ModelArk under the model ID dreamina-seedance-2-5-260628; the API is one way to use it, not the product itself. Is it open source? No. ByteDance has not published the weights, and same-named repositories on Hugging Face or GitHub are third-party wrappers.

Why people look for an uncensored Seedance

The model runs behind ByteDance's content policy: prompts, reference images and finished clips are reviewed, and requests that fail moderation are rejected. Creators making dance videos in revealing outfits, bedroom scenes and other adult themes often hit those refusals, so they search for an uncensored version. ByteDance offers no such version, and this page does not explain how to bypass its checks. Some third-party API resellers advertise looser Seedance access; they are not ByteDance products and we cannot confirm what filtering they apply. If you want uncensored generation, you can use Spicy AI: its image-to-video tool runs MiniMax H3 Turbo for consenting-adult content without content filters.

Seedance 2.5 vs Seedance 2.0: what changed

The upgrade is about length, references and control rather than raw resolution. The new model doubles a single take to 30 seconds, more than triples the reference pool, lets audio references stand alone and adds timestamp editing and Draft mode. The one thing it gives up is 4K, so teams that need 4K masters still use Seedance 2.0.

From the BytePlus ModelArk documentation for each model. See the separate Seedance 2.0 page for the older model.

Features: 30 seconds, 50 references, timestamp edits

A single request runs from 4 to 30 seconds (or -1 for automatic length), and a clip can be extended over several rounds into multi-minute pieces with smoother transitions between scenes. The reference pool holds up to 30 images, 10 videos and 10 audio files, used for characters, scenes, props, motion, clay-render looks and voices, and ByteDance says consistency across all of them improves. Timestamp-level editing lets you change the picture, camera, action or rhythm at a given second, and green-screen, camera and reference edits are built in. Output is 480p or 720p at 8-bit and 1080p at 10-bit, as MP4 or MOV. ByteDance itself says complex physical motion and multi-subject interaction still need improvement.

AI short dramas: one take per beat

Vertical micro-dramas are where the 30-second take matters most. With Seedance 2.0, studios built every scene from 15-second pieces and stitched them; the new model can hold a whole beat, such as a cold open or a confrontation between two speaking characters, in one continuous shot with dialogue and sound. Segmind published a test in which a single call produced a 30-second vertical micro-drama at 720x1280 with two characters and native audio, with no cuts or post-production. Episodes usually run longer than that, so teams still plan two or three takes per episode, keep characters stable with a shared reference sheet and use timestamp edits to fix one moment instead of regenerating the whole take. Dreamina also markets the model directly for short-drama production.

Draft mode: a quick 480p preview pass

Draft mode is new in this release. You first render a quick 480p draft to check composition and pacing, then generate the final clip at your target resolution from that draft. The two steps are billed separately, so treat the draft as a planning tool rather than a free preview, and never ship the draft itself as a final asset. It pays off most on long 20 to 30 second takes, where a wrong camera move is expensive to discover at 1080p.

Model ID and API parameters

Through BytePlus ModelArk, the model runs as an asynchronous task: you create a generation task, then poll it until the video is ready. The model ID is dreamina-seedance-2-5-260628. Core parameters are resolution (480p, 720p or 1080p), ratio (21:9, 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive), duration from 4 to 30 seconds or -1 for automatic, generate_audio, return_last_frame and the reference image, video and audio lists. The July launch post said API access was coming soon; the documentation is now live, though no single announcement gives an exact opening date. Requests rejected by moderation return an error instead of a video.

What creators make with it

The features ByteDance highlights map onto a handful of recurring creative formats. Thirty-second takes and multi-round extension suit short narrative pieces, while the 50-asset reference pool is what makes branded and character-driven series practical. The list below groups the formats we see most often in creator write-ups and the official showcase; it describes what the tools are used for, not a ranking of results.

• One-take short dramas: a 30-second scene with dialogue and a camera that follows two characters without a cut.
• Character series: the same face, outfit and voice across many clips, driven by a reference sheet of images and audio.
• Music and dance videos: choreography matched to an uploaded track, with multi-angle cuts inside one request.
• Product and e-commerce spots: a product photo, a lifestyle location and a voice-over line combined into one ad.
• Green-screen and camera re-edits: replace a background or change a camera move at a chosen timestamp instead of regenerating everything.
• Storyboard previz: Draft mode at 480p to test pacing before a final render.

How creators rate the new release

Reception so far is mixed in a predictable way. A roundup of Reddit threads by Leadde reports that creators praise the 30-second continuous scenes, dialogue and acting, character consistency and reference-driven generation, and that the recurring complaints are costly rerolls, weaker results in some action and dance scenes, moderation refusals and platform limits; some still prefer Seedance 2.0 as more forgiving. The predecessor set a high bar: Hedra notes that the previous model topped the Artificial Analysis video arena in early 2026. We found no public arena score for the model yet, so treat these as anecdotal signals and test on your own prompts.

Summarised from third-party write-ups (Leadde, Hedra); individual opinions, not lab benchmarks.

Seedance 2.5 prompt guide: shots, timestamps and references

Seedance follows structure better than long run-on prompts. Write each shot as its own line with framing, camera move, action, light and sound. On 2.5 you can address time directly, for example 0–5s for the opening wide shot and 5–10s for a close-up, and an edit can target one time range while keeping the rest continuous. Number every reference and say what it is for.

• Shot 1, Shot 2, Shot 3: one line each with shot size, camera, motion, light and audio.
• Timestamps such as 0–5s and 5–10s let an edit target one time range.
• Label references, e.g. image 1 = character, video 1 = camera move, audio 1 = voice; up to 50 are allowed.
• Use Draft mode at 480p to test framing and pacing before the final render.
• If identities drift or captions garble, reduce subjects, shorten shots and add references.

Prompt templates you can copy

Replace the bracketed parts and keep the structure. The timestamp templates use the 30-second length; the shot templates also work as a starting point in the H3 Turbo generator above, where your uploaded photo already defines the adult subject.

Step by step: a similar clip on Spicy AI

Spicy AI does not run this model, so this workflow uses MiniMax H3 Turbo, a different model, to get a similar short, prompted clip from your own photo. Expect a different look and shorter clips than the 30-second takes of the ByteDance model.

• 1. Open the image-to-video tool from the button at the top of this page; H3 Turbo is preselected.
• 2. Upload a clear photo of an adult subject that you own or have consent to use.
• 3. Paste a shot template from above and fill in action, camera, light and sound.
• 4. Generate a short clip first, check the motion, then change one element per retry.
• 5. For a longer story, render several short shots and join them in your editor.

Seedance 2.5 vs Wan 3.0 and MiniMax H3

Wan 3.0 also reaches 30 seconds at 1080p and accepts document and web input. Seedance adds a 50-asset reference pool, 10-bit 1080p and timestamp editing. MiniMax H3 is the opposite trade: open weights, 4 to 15 seconds with native stereo, and 2K through a re-render, while Seedance is closed and hosted. Pick on length, references, resolution and openness.

Prompt tips for Seedance-style results

1. Pick a photo with the framing you want in the first frame, because Spicy AI animates from an image.
2. Write shots in order ('Shot 1 … Shot 2 …'); results differ from Seedance because Spicy AI runs MiniMax H3 Turbo.
3. Name the dance style or camera move explicitly, such as K-pop choreography or a slow orbit.
4. Repeat outfit and lighting details in every shot to keep the look consistent.
5. Only use photos of adults (18+) you have the right to use, and never animate a real person without consent.

Sources

• Introducing the 30-second model (ByteDance Seed)
• Model documentation (BytePlus ModelArk)
• Seedance Reddit review roundup (Leadde, third party)
• Seedance overview (Hedra, third party)
• 30-second micro drama in one generation (Segmind test)
• AI short drama guide (Dreamina)

More uncensored AI Video models

• Unlimited Seedance 2.5
• uncensored seedance 2 0
• uncensored seedance 2 0 mini
• uncensored minimax h3
• Uncensored Kling
• Uncensored Kling 2.6
• Uncensored Kling O1
• Uncensored Sora
• Uncensored Veo
• Uncensored Hailuo AI
• Uncensored MiniMax H3 Max
• Uncensored Wan
• Uncensored Wan 3.0
• Uncensored Seedance
• Uncensored Seedance 2.0
• Uncensored Seedance 2.0 Mini
• Uncensored HunyuanVideo
• Uncensored LTX Video
• Uncensored Runway
• Uncensored Pika
• Uncensored Luma Dream Machine
• Uncensored Vidu
• Uncensored PixVerse
• Uncensored EdenLive
• Uncensored Grok Imagine