HiAPI
Blog

Guide

Seedance 2.5 Flash Image-to-Video: First Frame, Last Frame and Per-Second Billing

Making a still image move: parameters, price and the usual failure modes

Nov 17, 2026 7 min read
A discrete dot matrix in mint green on a black background, with dot density increasing gradually from left to right.
Model
seedance-2.5-flash
On this page

What this model does

seedance-2.5-flash/image-to-video is ByteDance's image-to-video model: you upload an image as the first frame and the model generates a continuous clip starting from it, with synchronized audio. The output aspect ratio follows the first frame rather than being re-cropped.

It fits work that starts from deterministic material: moving product shots, subtle expression and body motion for character art, dynamic previews of storyboard sketches, and filling a content pool with short clips in bulk. It is not for "one sentence in, a whole narrative out" — that is the text-to-video route's job.

Request parameters

  • prompt (required): the motion, camera move, scene and sound you expect, up to 5,000 characters.
  • first_frame_url (required): jpeg, png, webp, bmp, tiff, gif, heic or heif; under 30 MB; each side 300–6,000 px; aspect ratio 0.4–2.5.
  • last_frame_url (optional): same constraints as the first frame; a mismatched ratio is center-cropped to match. Supplying a last frame requires a first frame.
  • resolution: 480p or 720p, default 480p.
  • duration: an integer from 4 to 30, default 5, in seconds.
  • aspect_ratio: only adaptive — the output follows the first frame.
  • output_format: mp4 (default) or mov.
  • generate_audio: defaults to true.

Per-second billing

  • 480p: $0.0679/second
  • 720p: $0.1525/second

Billable seconds equal the duration you request. Common cases: 4 seconds at 480p is $0.2716, 4 seconds at 720p is $0.61, 10 seconds at 480p is $0.679, 10 seconds at 720p is $1.525. The first and last frame images themselves are not billed.

Because billing is per second, a retry is a second charge. Judging whether the first frame is right before you submit is far cheaper than regenerating.

The first frame sets the ceiling

The model can only extend what the first frame already contains, so its quality largely determines the result:

  • A clear subject occupying a sensible share of the frame — a tiny subject leaves the model unsure what should move.
  • Avoid heavy blur and noise; poor input gets amplified in every frame.
  • Watch the frame edges: a subject cut off at the border tends to deform once it moves.
  • Text inside the frame rarely stays stable in motion unless it is not the point.
  • Preparing several candidate first frames for the same subject beats endlessly rewording the prompt.

Duration is both a cost multiplier and a quality variable

The longer the clip, the more frames the model must keep consistent, and the higher the chance drift and distortion accumulate. Verify the first frame and the motion direction at 4–5 seconds, then extend as needed; build long pieces from several short clips edited together rather than relying on one long generation.

Common failures and the order to debug them

  1. A parameter error on submit. Check the image limits first (size, side length, aspect ratio, format) and whether you sent a last frame without a first frame.
  2. The subject deforms. Usually a small subject, a cut-off edge, or an oversized motion request. Change the frame or ask for less movement.
  3. The ratio is not what you expected. This route is fixed at adaptive and follows the first frame — set 16:9 or 9:16 in the frame itself.
  4. Content is blocked. The first frame goes through policy review as well; what is in the image decides whether the call can happen.
  5. The client times out but the task succeeded. Generation is slow; too short a client timeout turns success into failure, and resubmitting bills twice. Treat it as an asynchronous task with a generous timeout.

Recommendations before launch

Start with three to five different first frames and prove submit, poll and download end to end. Then hold the frame constant and vary only the prompt, to learn which motion descriptions the model responds to. Finally ramp in stages while watching success rate, latency and retry rate. The model page carries the current parameter reference and prices.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs