
Seedance 2.5 practical guide
Seedance 2.5 References: Image & Video Workflow
A reference helps when it has a clear role: appearance, starting composition, ending composition or motion. Start with the smallest set that describes the shot, then tell the prompt which role each file has. More files can introduce conflicting clothing, lighting and camera directions.
For the model overview, supported controls and a Studio with 2.5 selected, start at the Seedance 2.5 generator page.
Quick answer
In this Studio, use a first-frame image for the opening, an optional last-frame image for the ending and a video for motion. The customer API also accepts image/video reference lists subject to account limits. Audio-file reference uploads are not offered here; generated audio is a separate option.
Model capabilities and the controls available here
ByteDance describes broader multimodal reference capabilities in its Seedance 2.5 announcement. Those model capabilities are separate from a particular website's integration. Use the fields actually exposed in our Studio or API; an audio reference mentioned in an external demo is not an upload field on this service.
The web Studio exposes first-frame image, last-frame image and video-reference inputs. For an ending image, provide a first-frame image too and describe the physical transition between them. The API supports account-bound reference lists and frame fields; follow its current request limits rather than a model-wide reference count.
Give each reference one job
- First frame: establish the subject, product appearance, wardrobe, setting and starting composition.
- Last frame: anchor the desired final composition; keep the subject and scene compatible with the opening image.
- Video: demonstrate motion or camera rhythm. State whether it controls movement only, so its actor or setting does not compete with your image.
- API image references: assign a clear subject or scene role in the prompt, and avoid mutually inconsistent assets.
- Generated audio: describe ambience or sound in the prompt and enable Audio if wanted. This does not copy a supplied voice or soundtrack.
A Studio-ready reference prompt
Upload a first-frame image and, if useful, a short motion video. Add a compatible last-frame image only when the ending matters. The example refers to these named Studio inputs and does not require an unsupported audio upload.
Use the first-frame image for the person's face, navy coat, brown bag and cafe layout.
Use the video reference only for a slow walking camera rhythm; do not copy its actor, clothing or location.
Create one continuous cafe scene: the person walks from the entrance to the window table, sets down the bag and opens a notebook.
Preserve the face, coat, bag and table geometry. Keep morning light consistent and avoid cuts.
If a last-frame image is supplied, finish in that composition with a physically plausible transition.
Generated audio: soft footsteps, chair movement and quiet cafe ambience. No subtitles or music.Prepare files before uploading
- Use a sharp, uncluttered image with the subject clearly visible. Remove irrelevant people and competing products.
- Keep first and last frames consistent in clothing, geometry, lighting and aspect ratio.
- Trim the motion clip to the action that matters. Do not add a video only to obtain a different billing rate.
- The current upload endpoints accept JPEG, PNG or WebP images up to 10 MiB and MP4 or MOV videos up to 20 MiB. Files must pass format and media checks; changing the filename does not make a file valid.
- A successful purchase, verified eligible account and enough credits are required. Upload frequency, capacity and unused-file quotas also apply.
How API upload drafts work
Create a draft with the model, duration, resolution and audio setting you plan to generate with. Upload image/video files against that draft, then submit its ID with the reference URLs returned by the upload endpoints. Arbitrary externally hosted reference URLs are not accepted by the customer API.
A draft belongs to your account and is bound to a generation task. If you change the generation settings after preparing the draft, create a matching draft instead of reusing incompatible uploads. Keep the API key on your server and follow the documentation's retry and idempotency rules.
24-hour reference lifetime
Reference files and their access links expire after 24 hours and cannot be accessed afterward. The draft's expiry is the deadline for its uploaded files, so uploading another file later does not restart that deadline. Submit while the draft is valid and keep your original assets in your own storage.
The uploaded references are private task assets, not a public archive or permanent external-link service. Generated results follow a separate lifecycle. Save the outputs you need and check account history for submitted tasks.
When the output drifts
First remove the weakest reference. If identity changes, use one clear appearance source. If the camera copies the wrong setting, state that the video controls motion only. If the end frame causes an implausible jump, simplify the transition or choose a closer ending composition.
Change one reference or instruction at a time. Review a shorter test before a 30-second run, and compare the new result against the same identity, movement and ending checks.
Plan your Seedance 2.5 workflow
Start at the Seedance 2.5 overview and generator for available settings, pricing and access rules.
Seedance 2.5 Prompt Guide: Templates for 15–30s Videos
Write a shot brief, assign timed beats and fix conflicting instructions.
Seedance 2.5 Pricing: 15 vs 30 Seconds & Credit Costs
Compare duration, resolution and reference-video costs before submitting.
Next step: generate or integrate
Review credit packs and plans or follow the customer API documentation for server-side keys, uploads and task polling. Keep API secrets on your server. Reference files and links become inaccessible after 24 hours.
Open Seedance 2.5 StudioModel source and demos: ByteDance's Seedance 2.5 announcement. Templates and workflow advice are editorial guidance from this independent platform; illustrations are not generated-video test results.