Video Model — text, image & multimodal-to-video
Seedance 2.5 is ByteDance's next-generation video model, succeeding Seedance 2.0 as its top contender alongside Veo and Kling. Its headline is native 30-second generation in a single pass (from 4–30 seconds).
Reference control is the other jump: up to 50 multimodal inputs per request, up from 12 — images, video, and audio, each assigned a job in the prompt. The same reference stack drives generation, editing, and extension. Audio is co-generated in the same generation as the visuals, so dialogue, music, and effects land in sync. Prompt adherence is roughly 20% better than 2.0.
Variants at a Glance
| Variant | Type | Modalities | Max Res. | Talking Points |
|---|---|---|---|---|
| 2.5 (new) | Video gen / edit / extend | T2V, I2V, first & last frame, multimodal refs (image / video / audio) | 480p draft → 1080p render (4K for Enterprise) | Native 30s single pass, no stitching; up to 50 reference inputs (30 img / 10 vid / 10 aud); timestamped video editing; forward, backward & bridge extension; native audio co-generation; ~20% better prompt adherence; Draft Mode — preview at 480p, then render the exact result in 1080p |
| 2.0 | Video gen / edit / extend | T2V, I2V (+ audio / video refs) | 4K | Caps at 12 reference inputs; shorter native duration ceiling; live on Artlist today |
Draft Mode
No more burning full-price 1080p credits on a generation you haven't seen yet. With Draft Mode, you can preview a Seedance 2.5 result at 480p first, then commit to 1080p from that exact draft — no re-entering inputs.
How to use it
- Generate a Seedance 2.5 video at 480p
- Check out the result
- Click the "Render to 1080p" button on the draft's asset card
- Get the exact same output, now in 1080p
Good to know
- Both the 480p draft and the 1080p render are billed at standard rates for each step — no price changes
- This replaces the standalone 480p option — 480p now only exists as a draft step on the way to 1080p
- The same inputs carry over automatically to the 1080p render, so what you validated is what you get
- The option to render a draft in 1080p expires 7 days after the draft is generated
- Draft Mode is scoped to Seedance 2.5 only, for now
- Enterprise customers can also render a draft in 4K (3840×2160). 4K is not available on other plans
Key Features
Generation Modes
- Text-to-Video (T2V): Generate video from a text prompt
- Image-to-Video (I2V): Text prompt plus a first-frame image
- Reference-to-Video (R2V): Upload up to 50 reference inputs (up to 30 images, 10 video clips, and 10 audio clips) to guide generation — use them to define subjects, background, style, camera movement, or soundtrack.
- First & Last Frame: Text prompt plus first-frame and last-frame images; if the two aspect ratios differ, the first frame wins and the last frame is auto-cropped
- Native 30s in one pass: A full 30-second clip is generated in a single pass — no stitching, no scene-cut splicing, no seams
- Multi-shot generation: Generate several shots with cuts between them from a single prompt
- Multimodal reference generation: Feed image, video, and audio references and specify in the prompt what each is for — subject, background, composition, style, camera movement, motion, dialogue, music, or SFX.
Editing & Extension
- Video editing: Replace subjects, add / remove / modify elements, or locally redraw regions — driven by reference materials plus the prompt
- Timestamped edits: Scope a change to a time window — "replace 0:02–0:05" — leaving the rest of the clip untouched
- Forward extension: Continue a clip on from its last frame, holding style, rhythm, and camera logic
- Backward extension: Generate the moment before a clip, using its first frame as the endpoint
- Seamless bridge: Take two clips and generate the missing transition between them
- Task-type intelligence: The model infers the task (generate / edit / extend) from the prompt. Duration semantics differ: generate and edit = total output; extend = the extended portion only
Smart Controls
- Adaptive aspect ratio: Industry-exclusive — T2V picks the best ratio from the prompt; first-frame tasks match the input image; multimodal reference picks from the first recognized media file (video over image)
- Smart Duration: The model can pick an appropriate length in the 2–30s range
- Prompt adherence: Roughly 20% better than 2.0 — fewer generations before a usable result
Technical Capabilities
| Duration | Any integer 4–30s per clip, generated natively in a single pass. Smart Duration (-1) auto-picks 4–15s. |
|---|---|
| Resolution & color | 480p draft preview; 720p direct generation; 1080p final render via Draft Mode (upgrade an existing 480p draft, same inputs, no re-entry). 24 FPS, 10-bit color depth. 4K (3840×2160, with audio) is available to Enterprise customers only. |
| Aspect ratios | 16:9, 9:16, 4:3, 3:4, 1:1, 21:9, or Adaptive (model chooses). If a specified ratio differs from an uploaded image's, the image is center-cropped. |
| Reference inputs | Up to 50 total = 30 images + 10 videos + 10 audio clips. Images: 0–30, up to 4K each. Videos: 0–10, up to 4K each, 2–30s per clip, total under 30s. Audio: 0–10, 2–30s per clip, total under 30s. Video and audio totals are counted separately. |
| Audio | Native — co-generated in the same latent space as visuals; audio references can drive dialogue, music, and SFX. |
| Spoken languages | English, Chinese, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese, Korean. |
Limitations
- Single clip is capped at 30s — long-form still needs to be assembled from multiple generations
- Reference video and reference audio are each capped at 30s total per request (each file 2–30s)
- If a specified aspect ratio differs from an uploaded image's, the image is center-cropped; in first & last frame mode a mismatched last frame is auto-cropped
- The ~20% prompt adherence gain is a vendor-stated figure, not something we've benchmarked
- 1080p output is only available by rendering an existing 480p draft — there's no direct one-shot 1080p generation
- A draft's 1080p render option expires 7 days after the draft is generated
- 4K output is available to Enterprise customers only
- Video editing can return a slightly shorter clip. Because of how Seedance 2.5 processes frames, an edited video can come out up to about 0.3 seconds shorter than the original, with a few frames dropped. This can show up as a small skip or stutter. ByteDance documents this as expected model behavior. See Prompting Tips for ways to keep the original length.
Prompting Tips
- State clearly in the prompt which reference materials to use and for what — subject vs style vs motion vs audio — the model follows your assignments
- With 50 reference slots available, be deliberate rather than exhaustive: a few well-labelled references beat a pile of unlabelled ones
- Use Adaptive aspect ratio for text-to-video when the target placement is unknown — the model picks the best ratio from the prompt
- Set an explicit duration when output length matters; Smart Duration (-1) decides for you
- For first & last frame generation, match the aspect ratios of both images — otherwise the last frame is auto-cropped to fit the first
- For edits, give a timestamp range to scope the change and add "keep everything else unchanged" so the model doesn't touch unrelated areas
- Describe the audio you want in the prompt — dialogue, music, and SFX are co-generated with the visuals, not added in a separate pass
- Use Draft Mode to lock in composition, motion, and prompt fidelity at 480p cost before rendering the final 1080p version
- To keep an edited video the same length as your original, try one of these:
- Use an 8n+1 frame count. When your clip has 8n+1 frames (for example 97, 121 or 145), ByteDance reports that the edited output matches the input length. At 24 FPS, a clip that is a whole number of seconds long already has a frame count divisible by 8, so adding one frame does it: a 5-second clip is 120 frames, so export it as 121 frames (5 seconds and 1 frame). Check the frame count in your video editor, add or trim a few frames, then upload that version. Step-by-step examples are in Troubleshooting AI Image and Video.
- Edit a clip generated by Seedance 2.5. If the clip you are editing was itself generated by Seedance 2.5, the edited output keeps its exact length. This does not apply to camera footage or clips from other tools.
- Use Reference-to-Video instead of edit. Use your video as a reference and set the duration yourself. You get exact control over length, but the result is less strictly tied to your original footage.