🔗 Seedance 2.5 Video Extension: Going Past 30 Seconds

Thirty seconds is where the duration field stops on seedance_2_5. We read it off Higgsfield's live model catalog on August 7, and no prompt phrasing moves it. The way past the wall is a mode, not a bigger number: video_extension takes a clip you already like and generates more of it, and ByteDance's own launch notes say a clip can be extended twice. Call it roughly 90 seconds of continuous footage before anyone reaches for a cut.

This guide covers the mode as it shipped, which chains are worth building, and the two things we cannot vouch for yet.

By the numbers#

Use the full 30 first#

Before chaining anything, spend the single pass. Thirty seconds in one generation is double what Seedance 2.0 allows. One pass means one grade, one motion arc, no joins to hide. A lot of "how do I extend this" problems are really "I generated 12 seconds when the field goes to 30" problems.

Here's an honest look at what a full-length 2.5 pass does and doesn't hold together:

When 30 genuinely isn't enough, the mode takes over.

How video_extension works on Higgsfield#

The job shape is simple. Set mode to video_extension, feed the finished clip as the video reference, and set extension_mode to forward or backward (the one field that's required only on these jobs). The duration field sets how much new footage you get, 4–30 seconds per pass. Whatever aspect you type is beside the point, because the output follows the source clip's aspect ratio. Resolution stays 480p or 720p, and your image, video and audio references can ride along to hold identity across the join.

One habit from our rig: QC the base clip completely before extending it. An extension inherits everything, including the mistake you hoped nobody would notice.

The habit itself predates the parameter. Creators were already running extension chains in the 2.0 era, and watching that process shows exactly what the 2.5 mode folds into a single field:

Forward, backward, and which one you want#

Forward continues past your final frame. That's the obvious one, and it's how a 30-second take becomes 60.

Backward is the sleeper. It generates footage that leads into your clip and ends where your clip begins, which is exactly what you want when the payoff came out perfect but there's no run-up: the product reveal with no approach shot, the punchline with no setup, the drop with no build. We'd also reach for it when a generation starts mid-action and needs two seconds of establishing air.

Both directions keep the source aspect, so a 9:16 short stays 9:16 down the whole chain. Set the aspect right on the base generation. That decision is permanent.

The 90-second math, honestly labeled#

ByteDance says 30 seconds in a single pass and extendable twice. 30 + 30 + 30 is about 90 seconds of continuous footage. Their two numbers, our arithmetic. The mode is days old on Higgsfield and we have not yet run a full two-extension chain on our account, so treat 90 as the ceiling to plan against, not a tested guarantee.

A chain is also three renders and two joins, and every join is a QC gate. Identity drift compounds across passes, so feed the same reference stack on every pass, exactly as you would across shots in a multi-shot sequence.

Extension or a cut?#

A cut costs no render and hides its join in plain sight.

If the format wants cuts anyway (trailers, listicles, anything fast-cut), stitch separate generations and keep 30-second passes as your longest unit. Extension earns its render when the shot must not cut: one-take POV pieces, dance coverage where the move crosses the boundary, walk-and-talks, a slow push-in that a cut would deflate. The difference on screen is continuity of motion. A camera move that carries through the join is the one thing a stitch cannot fake.

New to the model entirely? The complete guide covers the basics before you start chaining.

Audio across the join: the part we can't vouch for#

generate_audio defaults to true, so every extension pass comes back with native sound. What nobody has shown us is whether ambience and score carry across the boundary as one continuous bed, and we won't claim it until we've chained it ourselves.

Two safe plays in the meantime. Pass generate_audio: false on every pass and lay a single music bed across the finished chain in post; that's also the right call for shorts that will carry in-app trending audio. Or burn one throwaway chain as an audio test before promising a client continuous sound. Frame rate is unpublished too, which is worth knowing before you conform passes in an NLE.

The video_edit near-miss#

Here's the trap on long footage. Say second 22 of a 28-second master breaks: video_edit can repair the region, but it bills by the reference video's duration, and its duration and aspect fields are ignored. A small fix on a long master bills like the master. Overall credit pricing is unpublished, so we can't tell you what any of this costs in absolute terms. What we can tell you is the billing basis, and it punishes fixing late. Approve each pass before extending it; the deeper the chain, the more expensive a rescue. The editing features guide covers the repair suite itself.

How we checked#

Every parameter here was verified against Higgsfield's live model catalog on August 7, read off model id seedance_2_5 on our own account. No exceptions. The 30-second single pass, the two-extension limit and the July 31 consumer launch through Jimeng and Doubao are ByteDance Seed's statements. The Higgsfield model page carries the marketing framing; the catalog carries the shipped fields, and where they disagree we report the catalog. We have not yet rendered a two-extension chain, and we say so above rather than pretend.

Last verified: August 7, 2026.

The 30-second wall was the defining constraint of AI video for two years. On this catalog it's a routing decision, and the interesting ceiling is now 90.