Run ByteDance's Seedance 2.5 online: one continuous 30-second video in a single pass, up to 50 reference files, and audio that drives pacing and lip-sync. Test it here — drafts stay free on the fast Seedance tiers.
ByteDance's new flagship breaks the ceiling every video model shares: instead of 5-to-15-second fragments, Seedance 2.5 generates one continuous 30-second shot in a single pass — no stitching, no seams. Feed it up to 50 reference files, or hand it a music track and let the audio drive the pacing, the cuts and the lip-sync. The Seedance you've been drafting with here just got a big sibling.
Draft free on Seedance fast tiers · credits only on the long take
The model finished rolling out days ago, so these are early impressions from our test renders plus ByteDance's launch numbers — flagged as such. Independent benchmarks still describe its predecessor; we'll revise as they land.
• The 30-second single take.
One native generation, not stitched clips — so the camera move, the character and the light hold across the full half-minute. Multi-turn extension goes further, and a 180-second Ultra-Long mode exists in beta upstream.
• 50-file reference ceiling.
Up to 30 images, 10 video clips and 10 audio clips per generation — the largest reference capacity of anything we host, built for teams with a real asset kit.
• Audio drives the video.
Supply one music, voice or SFX track and it sets the pacing, beat-matches the cuts and syncs the lips — in 10+ languages. This inverts the usual workflow: the soundtrack directs the shoot.
• Director-grade controls.
Region-level edits that change part of a frame without regenerating the clip, green-screen background swaps that hold the subject, and white-model blocking for previz. These are production tools, not toy sliders.
• The 4K asterisk.
Native 4K was announced the same day — for the Seedance 2.0 pipeline. Some coverage folded it into 2.5's spec sheet; ByteDance's own three pillars for 2.5 are duration, references and control. Don't pick it for a resolution headline.
• Benchmarks pending.
Public arenas still score Seedance 2.0. Every "new king" claim this month — about this model or its rivals — is provisional until independent numbers land.
• No free lunch upstream.
On ByteDance's own platform, 2.5 launched with no free quota at all — enabling it requires a funded account. Anyone advertising "free unlimited Seedance 2.5" is reselling something, or fibbing.
• Long takes, long waits.
Thirty seconds of native video is a big render. Expect multi-minute generation times and plan credits accordingly — draft the idea on the fast tiers first, always.
Everything here exercises what only Seedance 2.5 can do in our lineup: length, audio-driven timing, and surgical edits. Copy, swap the brackets, render.
"One continuous handheld shot: we follow a [courier] from behind as she weaves through a crowded night market, ducks under a stall awning as rain starts, and hands a package to a vendor who grins. The camera never cuts. Neon reflections in puddles, escalating rain sound, market chatter throughout."
"The camera never cuts" is now an instruction you can actually give. Write the shot as one sentence of continuous action with a beginning, middle and end — the model holds identity and lighting because it's one generation, not three stitched ones.
Upload: a 30-second music track
"A sneaker ad cut to this track: product spins and street-run shots that change on the beat, slow-motion on the bass drops, final logo lockup on the last hit. Urban dusk palette, 24 fps."
The uploaded track isn't background music — it's the edit's timing authority. Naming what happens on beats and drops ("change on the beat", "slow-motion on the drops") tells the model how to spend its 30 seconds. No other model here takes direction from a soundtrack.
Start from: a clip you already generated
"In the left third of the frame only: replace the parked van with a flower stall. Everything else — the walking couple, the camera drift, the audio — stays exactly as it is."
Region-level editing means one wrong object no longer costs you a re-roll of a 30-second render. Fence off the region, name the swap, and explicitly protect the rest — the protection clause is what keeps the edit from bleeding.
The two newcomers shipped within days of each other, with Veo as the established yardstick. Same prompts across all three, judged by what we'd ship — provisionally, given how new two of them are. Compare them yourself from one prompt box.
previewed at Volcano Engine's FORCE conference on June 23, enterprise beta first, public rollout completing in early August. Native 30-second generation, 50 references, audio-driven sync. ByteDance skipped 2.1–2.4 to mark the jump. The version on this page.
the reigning arena champion for text-to-video and image-to-video through early 2026, with native audio and multi-shot consistency. The Fast tier is what powers your free daily drafts on this site.
the speed-and-price era that made Seedance the community's default sketchpad: unglamorous, fast, everywhere. That reputation is why "draft on Seedance" is still the standing advice on every page of this site.
ByteDance's new flagship video model, previewed at the Volcano Engine FORCE conference on June 23, 2026 and rolled out through late July and early August. Three headline upgrades: one continuous 30-second clip in a single pass, up to 50 multimodal references (30 images, 10 videos, 10 audio clips), and finer production controls like region-level editing. The version number jumps straight from 2.0 to 2.5 — ByteDance skipped 2.1 through 2.4 to signal a generational leap.
The free sketchpad. Draft every idea here before spending credits on the flagship.
The other newcomer. Cheapest 2K clips and instruction-based editing.
The realist. When the clip has to pass as actual footage.