What is SeeDance 2.5?
SeeDance 2.5 is ByteDance's next-generation video model, announced June 23, 2026 at the Volcano Engine FORCE conference in Beijing. It is not yet publicly callable: the model is in closed enterprise beta, with public access expected to start in ByteDance's own apps and a third-party API to follow in late July 2026. Every spec on this page is from the announcement and will be re-verified against the live API the day it opens. We will make it available on Unifically at API launch.
The headline: one continuous 30-second video per generation, native 4K output with 10-bit color, and up to 50 multimodal references in a single call.
What's new in SeeDance 2.5
Native 30-second generation
One pass produces a continuous 30-second video with scene changes and tempo shifts inside it. No stitching. SeeDance 2.0 tops out at 15 seconds per generation; most competing models sit between 8 and 15.
Up to 50 multimodal references
Images, videos, audio, style references, and low-fidelity 3D white models in one call. SeeDance 2.0 caps at 15 mixed inputs (9 images, 3 videos, 3 audio). The 3D white-model input is new: a rough blockout can drive motion and spatial layout.
Native 4K with 10-bit color
Rendered at 4K rather than upscaled to it, with 10-bit color depth for gradients and grading headroom. SeeDance 2.0 reaches 4K only on its Pro variant; on 2.5 it is the standard render.
Region editing
Redraw one element of a finished video, like clothing, a background, or a product, while performance and lighting stay untouched. Removes the regenerate-everything loop for small changes.
Joint audio-video generation
Audio is co-generated with video in a shared latent space, so lip-sync and effects timing come from the model rather than a post-process. Multilingual output, including English, Chinese, Spanish, Japanese, and Korean.
ByteDance also reports about 20% better prompt adherence than SeeDance 2.0, and lists a beta long-video mode that extends output to 180 seconds. Both are the lab's own numbers; no independent benchmark or Arena result exists yet, and we will add them once the model is public and tested.
Best for
Continuous 30-second narratives
One pass covers a full short-form story with scene changes inside it, instead of stitching 10-second segments and hiding the seams.
Reference-driven brand work
A 3D blockout, a style set, and an audio cue can all steer the same generation. Up to 50 mixed references per call.
4K masters for grading
Native 4K with 10-bit color holds up in a color pipeline. No upscale pass.
Iteration without regenerates
Region editing swaps wardrobe, backgrounds, or products in a finished video while the performance stays locked.
Dialogue-led videos
Audio generates in the same pass and stays in sync, with multilingual output including English, Chinese, Spanish, Japanese, and Korean.
SeeDance 2.5 vs SeeDance 2.0
SeeDance 2.0 is live on Unifically today, with Pro, Mini, Fast, and Low Restriction variants. It generates 4 to 15 second videos from 480p up to 4K on Pro (Mini and Fast cap at 720p) and takes up to 9 images, 3 videos, and 3 audio references per call. SeeDance 2.5 doubles the native duration to 30 seconds, multiplies the reference budget to 50, makes 10-bit 4K the native render, and adds region editing, which 2.0 has no equivalent for. Until 2.5 is callable, 2.0 stays the practical pick; if your work depends on longer single-pass videos or 4K masters, it is worth waiting for the 2.5 API before committing a pipeline.
Availability
Announced June 23, 2026. Closed enterprise beta as of early July 2026. ByteDance's stated rollout order: its own consumer apps first, then CapCut, then third-party API access in late July 2026. No public pricing, variant list, aspect-ratio set, or per-type reference limits have been published. This page will be updated with exact callable values, live pricing, and Arena results as soon as the model is out.
FAQs
People also ask
SeeDance 2.5 is ByteDance's next video generation model, announced on June 23, 2026 at the Volcano Engine FORCE conference. It generates one continuous video up to 30 seconds with native audio, accepts up to 50 multimodal references in a call, and outputs native 4K with 10-bit color. It is not yet publicly available.
ByteDance targeted early July 2026 for public access, starting in its own apps, with third-party API access expected around late July 2026. It is in closed enterprise beta right now. We will open it on Unifically as soon as the API is live.
The native duration ceiling doubles from 15 to 30 seconds in a single pass, the reference limit grows from 15 mixed inputs to 50, 4K with 10-bit color becomes the native render (SeeDance 2.0 reaches 4K only on its Pro variant), and a new region-editing mode redraws one element of a finished video without regenerating the rest.
Yes. Audio is generated jointly with the video in the same pass, in a shared latent space, so dialogue and effects stay in sync. ByteDance lists multilingual audio support including English, Chinese, Spanish, Japanese, and Korean.
Up to 50 mixed inputs per call, spanning images, video, audio, style references, and low-fidelity 3D white models that guide motion and spatial layout. ByteDance has not yet published the per-type breakdown.
You can redraw one element of a generated video, like clothing, a background, or a product, while the rest of the video, including performance and lighting, stays untouched. SeeDance 2.0 has no equivalent; a change meant a full regenerate.
30 seconds in one continuous generation, with scene changes and tempo shifts inside that single pass. ByteDance also lists a beta long-video mode that extends output to 180 seconds.
Yes. We add SeeDance models at API launch. Pricing, variants, and exact parameters will appear here once ByteDance publishes them.
