Announced at ByteDance's Volcano Engine FORCE conference in June 2026, Seedance 2.5 is the first major AI video foundation model to generate a continuous 30-second clip at native 4K without stitching segments. Combine up to 50 multimodal references, edit regions inside a frame, and co-generate synchronized audio — then try the Seedances AI Video Studio below.
Text-to-video, image-to-video, and reference-to-video — generate up to 30s with Seedance 2.5.
Latest — up to 30s, 480p/720p, richer reference inputs
Tip: Press Ctrl+Enter to generate. Put dialogue in quotes for lip-synced audio.
0 / 5000
Sign in to start generating AI videos with Seedance 2.5.
Your preview will appear here
Write a prompt, choose settings, and hit Generate — your cinematic clip renders in this panel.
ByteDance positions Seedance 2.5 as a production-grade upgrade over Seedance 2.0 — longer native clips, richer reference control, and editing tools that match how real teams ship video.
Earlier AI video models stitched shorter segments, which caused character drift, lighting shifts, and visible seams at cuts. Seedance 2.5 uses a sparse diffusion transformer to maintain scene state across the full timeline in one inference pass — scene changes and tempo shifts included.
Seedance 2.5 targets 3840×2160 output generated natively rather than upscaled from a draft. Ten-bit color depth preserves smoother gradients for brand film, product hero shots, and post-production grading.
Lock character identity, product look, camera style, and sound design by attaching images, video clips, audio beds, 3D assets, and style guides in a single job — a major jump from the dozen inputs supported on Seedance 2.0.
Select part of a frame — a product label, background, wardrobe element, or prop — and regenerate only that region while preserving motion, camera path, and lighting everywhere else. Fix mistakes without rebuilding the entire 30-second clip.
Visual and audio signals share a unified latent space, so dialogue, ambient sound, and effects stay synchronized with on-screen action. Wrap spoken lines in double quotes for clearer lip-sync timing.
Built by ByteDance's Doubao team, the sparse attention architecture extends temporal context so the model tracks identity, physics, and camera continuity across long clips — the technical foundation behind single-pass 30-second output.
Whether you start from a script, a product photo, or a full reference board, Seedance 2.5 supports the same production routes teams already use on Seedance 2.0 — with room to scale references and duration.
Describe subject, environment, camera verbs, pacing, and mood in natural language. Seedance 2.5 is tuned for longer narrative beats — up to 30 seconds with internal scene transitions — and ByteDance reports roughly 20% stronger prompt adherence versus the previous generation.
Upload hero frames or first-and-last-frame pairs to animate catalog photography, key art, and storyboards. Image-to-video keeps labels, materials, and composition faithful while Seedance 2.5 adds camera motion, physics, and native audio in the same render.
Attach up to 50 references — portraits, product shots, motion plates, music cues, and style frames — then describe how each asset should be used. Seedance 2.5 maps multimodal inputs to characters, wardrobe, camera behavior, and sound design across the full clip.
Seedance 2.5 is ByteDance's next-generation AI video model, previewed at the Volcano Engine FORCE conference and rolling out through enterprise and consumer channels in mid-2026. It targets teams who outgrew 10–15 second clip generators and manual stitching workflows.
For most of 2024 and 2025, AI video meant short loops: five seconds here, ten seconds there, then an editor hid the seams. Seedance 2.5 attacks the duration problem at the architecture level. ByteDance claims the model produces an unbroken 30-second timeline in a single generation call — including deliberate scene changes and tempo shifts — without post-stitching shorter segments. That matters for performance marketing hooks, vertical stories, education explainers, and pre-visualization where narrative pacing cannot break every twelve seconds.
Resolution and color science move forward as well. Where Seedance 2.0 topped out at 1080p for many routes and added 4K later in the family, Seedance 2.5 advertises native 4K (3840×2160) with 10-bit color depth. Gradients in skies, skin tones, and product surfaces survive grading better than 8-bit drafts. Teams can preview faster at 720p inside Seedances, then export at 4K for delivery — a workflow that mirrors traditional post houses without leaving the studio.
Reference capacity scales for brand-consistent output. Seedance 2.0 reference mode supports roughly a dozen multimodal inputs — nine images plus limited video and audio. Seedance 2.5 raises the ceiling to about 50 simultaneous references spanning images, video clips, audio, 3D assets, and style guides. That enables multi-character scenes, strict product compliance, and campaign systems where every variant must match approved art direction across a full half-minute clip.
Post-generation control is the other headline. Region-level editing lets creators mask part of a frame — swap a label, replace a background, adjust wardrobe — while motion, camera trajectory, and lighting outside the mask remain untouched. Combined with clip-level storyline edits ByteDance demonstrated on Seedance 2.0, Seedance 2.5 moves AI video closer to iterative finishing instead of all-or-nothing re-rolls that burn credits and time.
Seedances wraps Seedance 2.5 in a production-ready web stack: authenticated accounts, credit-based billing with automatic refunds on failure, subscription tiers, and a unified generation history. Generate Seedance 2.5 now in the AI Video Studio — text, image, and reference routes with native audio — or continue with the Seedance 2.0 family on /create.
Seedance 2.5: native 30s single-pass video · native 4K · 50 multimodal references · region editing · unified audio · Sparse Diffusion Transformer · studio at seedances.ai/seedance-2-5.
Longer native clips and richer controls unlock workflows that short-loop generators could not support without heavy manual editing.
Ship 30-second paid social spots, seasonal campaigns, and spokesperson variants from one prompt board. Region editing localizes labels and offers without reshooting. Multimodal references keep talent wardrobe and product packaging consistent across dozens of variants.
Turn studio photography into lifestyle scenes, tabletop reveals, and 360-style motion without physical sets. Native 4K preserves label legibility for marketplaces that reject soft upscales. Batch hooks at 720p, then export winners at full resolution.
Build half-minute concept explainers with synchronized narration — dialogue in quotes drives lip-sync on presenter-style shots. Scene changes inside one clip support chapter-like pacing for LMS modules and YouTube explainers.
Directors block camera moves and lighting across 30-second beats before live action. Agencies pitch storyboards as motion with native audio temp tracks. Sparse attention helps maintain character continuity across longer beats than earlier models allowed.
Common questions about Seedance 2.5 and generating on Seedances.
Seedance 2.5 is live on Seedances. Generate longer native clips with text, image, or reference inputs in the AI Video Studio.
Open AI Video Studio
