The spec sheet for Seedance 2.5 reads like a wishlist: takes up to thirty seconds, up to fifty reference images, reference video input, a generated soundtrack, six aspect ratios. Specs are easy to print and hard to trust, so this review is organized around a different question — what does each number buy you on a real job?
Every figure below comes from the model registry Imgo routes against — the same numbers the API validates — so nothing here can quietly drift from what you can actually submit.
Seedance 2.5 is the model you reach for when the shot runs long or carries evidence: a character that must stay recognisable across scenes, a product that appears from three angles in one take, a vertical and a square cut from the same source. It is not the model for a take that must re-render identically — that is Veo 3.1's seed, and no amount of references replaces it.
The registry lists 5 / 10 / 15 / 20 / 25 / 30-second durations. In practice the long end changes how you plan: a 25-second product film stops being "four stitched eight-second shots" and becomes one request with one lighting continuity. That is the difference between a video that feels assembled and one that feels filmed.
The honest caveat: longer takes are more prompt-sensitive. At five seconds, a vague instruction still lands because there is no time to drift. At thirty, drift has room — lock the shot list in the prompt ("[beat 1]… then [beat 2]… closing on [beat 3]") or the extra seconds spend themselves on hesitation.
Fifty reference-image slots sounds like a party trick. It is actually a consistency engine. The working patterns:
What fifty slots are not is fifty unrelated images. Every additional reference dilutes the others; eight purposeful references beat fifty hopeful ones on every run we made.
Reference-video input is the quiet superpower: feed a fifteen-second clip of real motion — pouring, walking, fabric moving — and the generated take inherits the timing. Audio generation means the clip ships with sound; turn it off when a voiceover stage exists downstream, or you pay for audio you discard.
| Seedance 2.5 | Veo 3.1 | |
|---|---|---|
| Clip length | 5–30s | 4 / 6 / 8s |
| Reference images | up to 50 | up to 3 |
| Reference video | yes | no |
| Reproducible seed | no | yes |
| Resolutions | 720p / 1080p / 4K | 720p / 1080p / 4K |
Two constraints worth repeating from the registry: Veo renders eight seconds whenever output is 1080p or 4K or starts from a frame; and an end frame only travels with a start frame. Those, plus the seed, are why approved-shot pipelines still land on Veo. The full breakdown is in the Seedance 2.5 vs Veo 3.1 comparison, and the post-Sora landscape — the API closed September 24, 2026 — is mapped in the three-way comparison.
The strongest pipeline we settled on is not one model: make the first frame in an image model (Nano Banana 2 for speed, Nano Banana Pro when the character must hold), feed it to Seedance as the start frame with the character references attached, and keep Veo for the eight-second hero insert that has to re-render for the client on Thursday. One generator, one credit balance, model as a field.
If your videos are single-shot, eight seconds or under, and never need to match a previous take — Veo's seed makes it the safer default, and Kling 3.0's frame-locked control covers the short reproducible case at a different price point. Seedance 2.5 earns its credits the moment length, evidence, or multi-angle consistency enters the brief.
Prices move; the live credit quote in the generator before you submit is the only number that counts. The Seedance 2.5 model page carries the current routing status.
No image
Seedance 2.5's headline numbers — thirty seconds, fifty references — are registry facts. What they buy you in practice, and where Veo 3.1 still takes the job, is what this review is for.
No image
H3's weights are on Hugging Face, but H3 is a three-module system and only H3-Base shipped. Here is what runs locally, why 2K and the context orchestrator stay hosted, what four GPUs get you, and the territory clause most coverage skipped.
No image
OpenAI switches off the Sora API on September 24, 2026 — the app already closed in April. Here is what actually ends, how Sora 2's capabilities map onto Seedance 2.5 and Veo 3.1, and a checklist for migrating this week.