Seedance 2.5

Overview

Seedance 2.5 is ByteDance Seed's video generation model, launched on July 31, 2026. ByteDance calls it "the new-generation video creation model" and leads on one capability: length in a single shot. The launch post's core claim is that it can "generate high-quality, 30-second audio-video clips in a single pass and supports multiple rounds of extension."

That inherits the architecture that defined its predecessor. Seedance 2.0 introduced a unified multimodal audio-video joint generation architecture — video and audio produced together by one model rather than dubbed afterwards, across four input modalities. Seedance 2.5 keeps that and scales two dimensions that creators actually hit limits on:

Reference capacity is the more consequential of the two. It is what lets a creator pin a cast, a set and a soundtrack simultaneously rather than re-describing them in a prompt and hoping for consistency. ByteDance frames the target as narrative structure — building "setup, development, turning points, and resolution" inside one generation.

Seedance is the video half of ByteDance's generative media stack; Seedream 5.0 is the image half. For the launch itself, see Seedance 2.5 Ships: 30-Second Video in a Single Pass.

What this page can and cannot source

Seedance 2.5 shipped without the documentation its predecessor had. There is no technical report for 2.5, no stated output resolution in any engineering source, no published rate card and no independent benchmark. Seedance 2.0 has all four.

This page therefore separates three tiers of claim, and labels them throughout:

1. ByteDance's engineering blog — the July 31 launch post and the Seed product page. Vendor claims, but the vendor's technical voice.
2. ByteDance's consumer marketing — the Dreamina page, which advertises figures the engineering blog does not state. First-party, but marketing. This split is not new: for Seedance 2.0, the technical report specified 480p and 720p as native resolutions while the consumer page advertised "720p–1080p."
3. Independent measurement — arena.ai and Artificial Analysis. For 2.5 this tier is empty.

A note on naming

Seedance 2.5 was first announced on June 23, 2026 at the Volcano Engine FORCE conference in Beijing (Volcano Engine's event page confirms the 2026.06.23–06.24 dates, though it names no model versions), and then did not exist as a product for five weeks. As late as July 8, 2026 ByteDance's own Dreamina page read "Seedance 2.5 is coming soon!" while a wave of third-party posts described the model as released, shipped or "out now." Capability figures circulated during that window that no primary source supported.

The launch is now confirmed on three primary checks, read on August 2, 2026: the Seed tech blog carries the launch post; ByteDance Seed's model directory lists Seedance 2.5 above Seedance 2.0, 1.5 pro and 1.0; and the Dreamina page reads "New model LIVE: Dreamina Seedance 2.5 is now available."

Capabilities

From ByteDance Seed's launch post unless marked otherwise. These are vendor claims; no independent evaluation exists.

• 30-second single-pass generation — audio and video together in one pass, "with multiple rounds of extension" for longer pieces. The Seed product page is more specific about the ceiling: "Create videos up to 30 seconds in a single generation, with the option to extend twice."
• Continuity across cuts — the model "improves shot transitions and scene changes for stronger continuity in longer videos," and ByteDance claims consistency of main characters, environments and narrative pacing holds across extension rounds.
• 50 reference items in one pass — "up to 30 images, 10 video clips, and 10 audio clips as reference materials." ByteDance also says it strengthens "clay render, motion, and creative references."
• Timestamp-level editing — the model "offers timestamp-level control for targeted editing of audio and video content," changing what happens at a specific moment rather than regenerating the clip.
• Green screen editing — "the model can replace backgrounds and tell entirely different stories while keeping the main subject intact," with physical responses rendered: "the fluttering direction of clothes, the state of hair, gait rhythm, and lighting interaction."
• Camera perspective and reference-based editing — both named as enhanced over 2.0. The product page adds white-model control, a previsualisation workflow.
• Fewer uncontrolled artifacts — ByteDance claims it minimises "uncontrolled occurrences in subtitles and background music," two of the tells that mark AI-generated footage.
• Narrative structure — generating "setup, development, turning points, and resolution" within a single pass.

Consumer-page additions (marketing tier). Dreamina advertises local editing — "Edit specific video regions without regeneration" while "preserving lighting, composition, motion, audio, and timeline continuity" — a beta long-video mode up to 180 seconds, and prompt support in 11 languages including Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean. The engineering blog states none of these.

Technical Specifications

• Developer: ByteDance Seed
• Architecture: unified multimodal audio-video joint generation, inherited from Seedance 2.0. ByteDance published no architecture detail with 2.5; for 2.0 it credited "the efficiency of a sparse architecture."
• Parameter count: not published by ByteDance, for either version. This page does not state one.
• Input modalities: text, image, audio, video
• Output: video with synchronised audio
• Duration: up to 30 seconds in a single generation, extendable twice (Seed product page)
• Reference input limits: up to 30 images, 10 video clips, 10 audio clips per generation — 50 items
• Native output resolution: not stated in any ByteDance engineering source. Dreamina's consumer page advertises 4K; no technical report or rate card corroborates it
• Aspect ratios: not published for 2.5
• Model ID: no Seedance 2.5 ID is visible in Volcano Engine's or BytePlus's documentation. Aggregator sites quote doubao-seedance-2-5-260628; treat as unverified
• Weights: closed. No Hugging Face checkpoint, no self-hosting, no fine-tuning
• Technical report: none published for 2.5. Seedance 2.0's is arXiv 2604.14148, April 15, 2026

Seedance 2.0 specifications, for comparison

The predecessor is documented far more precisely, and its figures are the only ones on this family that come from a peer-facing paper:

• Native output resolutions: 480p and 720p (technical report). Dreamina advertised "720p–1080p" and BytePlus published a 1080p tier; the paper does not describe 1080p as native, and the arena.ai entry that topped the image-to-video board is explicitly the 720p build
• Duration: 4–15 seconds
• Reference input limits: 3 video clips, 9 images, 3 audio clips
• Aspect ratios: 6 (per CapCut's newsroom, unenumerated)
• Volcano Engine model ID: doubao-seedance-2-0-260128; doubao-seedance-2-0-fast-260128 for the Fast variant
• arena.ai model string: dreamina-seedance-2.0-720p
• Variants: Seedance 2.0 and Seedance 2.0 Fast (accelerated, low-latency); Dreamina additionally marketed a Seedance 2.0 Mini tier that the technical report does not describe

Use Cases

• Short-form narrative and micro-drama — the 30-second single-pass ceiling covers a complete beat rather than a shot, which is the specific gap 2.0's 15 seconds left open.
• Reference-driven ad creative — supply product stills, a reference clip and a voice track and generate variants against them. The jump to 50 reference items is what makes a full brand kit expressible in one pass.
• Targeted editing of existing footage — timestamp-level control for fixing a moment rather than regenerating a clip.
• Background replacement and compositing — green-screen editing that keeps the subject and its physical responses intact while the setting changes.
• Previsualisation — white-model control and camera-perspective editing, for blocking a shot before committing budget.
• Education — ByteDance names lesson material, creating "vivid and immersive visuals" for teaching.
• Synthetic training data for robotics — ByteDance names manufacturing, generating "synthetic video data" for robot training.
• Autonomous driving simulation — ByteDance names simulating "long-tail scenarios, such as extreme weather and complex traffic conditions."

The last three are ByteDance's own industry framing in the launch post, and it is a notable shift: 2.0 was sold to creators, while 2.5 is also pitched as a data generator for machine consumption.

Performance / Benchmarks

No independent benchmark covers Seedance 2.5. As of August 2, 2026 neither arena.ai nor Artificial Analysis had scored it, and ByteDance published no benchmark chart with the launch — not even the internal SeedVideoBench radar charts it used for 2.0. There is nothing to report for this version, and nothing on this page should be read as a measurement of it.

The Seedance 2.0 record

The independent numbers below are Seedance 2.0's. They are retained because they establish where the family stood before 2.5 and because they are the only third-party measurement of any Seedance model.

These come from two organisations with two different Elo scales — arena.ai (the former LMArena) and Artificial Analysis. The scores are not comparable across boards. Figures were read live on July 8, 2026; leaderboards move.

arena.ai (model string dreamina-seedance-2.0-720p)

On text-to-video, Google's gemini-omni-flash led at 1527 ±13 — on only 5,449 votes, which arena.ai flags as preliminary, against Seedance's 41,953.

Artificial Analysis (model string "Dreamina Seedance 2.0 720p")

On the no-audio text-to-video board, Alibaba-ATH's HappyHorse-1.0 led at 1290.

ByteDance's own evaluation for 2.0 was SeedVideoBench-2.0, published as radar charts without printed scores. The 2.0 technical report's self-assessment was deliberately modest: the model "has demonstrated performance on par with the leading levels in the field."

The "#1 everywhere" claim

A claim circulating in mid-2026 holds that Chinese labs occupy #1 in every video arena category. On the boards checked here, that was false. Google's gemini-omni-flash was #1 on arena.ai text-to-video. The accurate statement is narrower and still notable: Chinese labs held #1 on all four Artificial Analysis video boards (Seedance 2.0 on three, HappyHorse-1.0 on the fourth) and on arena.ai image-to-video, with runners-up — HappyHorse, Wan, Kling — also Chinese.

Limitations

• Nothing about 2.5 has been independently verified. No benchmark, no technical report, no rate card, no stated resolution. Every capability figure in Overview and Capabilities is ByteDance's own, from a launch post.
• ByteDance concedes two failure modes. Its closing assessment: there is "still room for improvement, particularly regarding the physical plausibility of complex motions and the stability of scenes involving interactions among multiple subjects."
• Resolution is a marketing claim, not a documented one. 4K appears on the consumer page and nowhere in the engineering material. The same gap existed for 2.0, where the paper said 480p/720p and the consumer page said "720p–1080p."
• Unresolved copyright litigation risk, carried over and unaddressed. Following viral clips of recognisable actors and characters, The Hollywood Reporter reports that Disney, Paramount, Netflix, Warner Bros. and Sony each sent legal threats to ByteDance in February 2026. On February 20, 2026 the Motion Picture Association sent a cease-and-desist letter; The Hollywood Reporter writes that "it's this first time the MPA has sent a cease-and-desist to a major generative AI company." MPA chairman-CEO Charles Rivkin said, quoted by TheWrap: "In a single day, the Chinese AI service Seedance 2.0 has engaged in unauthorized use of U.S. copyrighted works on a massive scale… ByteDance should immediately cease its infringing activity." The MPA letter argued that "the scale and consistency of these results demonstrate systemic infringement rather than inadvertence." Disney's lawyers called the model a "virtual smash-and-grab of Disney's IP" (TechCrunch). The dispute remains unresolved, and the Seedance 2.5 launch post does not mention it — nor the safeguards ByteDance pledged, nor how availability outside China will be handled this time. A model accepting 50 user-supplied references per generation enlarges the question rather than closing it.
• 2.0's global rollout was paused over that dispute. TechCrunch, citing The Information, reported on March 15, 2026 that ByteDance "had planned to make Seedance 2.0 available globally in mid-March" but was "delaying those plans as its engineers and lawyers work to avert further legal issues." Availability outside China arrived later, region by region, with guardrails — CapCut's initial international release blocked "the ability to make videos from images or videos that contain real faces."
• No API yet. ByteDance says BytePlus ModelArk access is coming; BytePlus's product page still advertises Seedance 2.0. Anything built against 2.5 today runs through consumer surfaces.
• Closed weights. No self-hosting, no fine-tuning, no offline generation, and a dependency on ByteDance's platform terms. For open weights, Wan 2.2 is Apache 2.0.
• Chinese content policy. Generation is shaped by content rules that differ from those governing US-hosted models.

Pricing & Access

Seedance 2.5 — no published rates

ByteDance has published no rate card for Seedance 2.5. Volcano Engine's pricing documentation still covers the Seedance 2.0 series, and BytePlus ModelArk still lists Seedance 2.0. Aggregator sites quote per-token prices and a doubao-seedance-2-5-260628 model ID; those figures have no primary source and this page does not repeat them as fact.

Access at launch is through consumer surfaces:

• Jimeng AI — the Chinese creative platform, named in the launch post as a rollout surface
• Doubao Pro — ByteDance's assistant, on the paid tier
• Dreamina — the international Jimeng surface, which reports the model live with free daily credits and paid subscription plans; it publishes no per-generation rate
• BytePlus ModelArk — the international API, announced as coming, undated

Seedance 2.0 — the published rates

These remain the only rates ByteDance has published for the Seedance family. Announced March 4, 2026, billed by token count derived from resolution, frame rate and duration — not per second.

Worked example, as reported by ITHome/Sina and BAAI's community hub: a 15-second 720p/24fps clip consumes roughly 308,880 tokens — about 8.65 元 in editing mode and about 14.21 元 for pure generation. The latter is a derived figure and the origin of the "1 元 1 秒" framing; ByteDance does not bill per second. These are CNY prices and this page does not convert them.

Seedance 2.0 also remains reachable through CapCut (Jianying in China), where TechCrunch dates the international rollout to March 26, 2026, expanding to Europe (March 28), Japan (March 31) and the United States (April 7, 2026); and through Volcano Engine / Volcano Ark, the China API, reported "全面开放" (fully open) on April 14, 2026.

There are no open weights and no self-hosted deployment option for any Seedance version. Rates change; confirm against Volcano Engine's model pricing page before budgeting.

Ecosystem & Tools

• Seedance 2.5 product page — ByteDance Seed's official page for the current model
• Dreamina — ByteDance's consumer creative platform (Jimeng in China), the first surface to carry 2.5
• Seedance 2.0 product page — the predecessor's page, with the SeedVideoBench-2.0 charts
• Volcano Ark — the China model platform, experience centre and API
• BytePlus ModelArk — the international API surface, currently advertising Seedance 2.0
• CapCut / Jianying — the video editor the Seedance line is embedded in; The Batch puts CapCut at 736 million monthly active users
• Seedream 5.0 — the sibling image generation line, same team, same Volcano Engine stack
• Seedance 2.0, 1.5 pro and 1.0 — the prior generations, all still listed in ByteDance Seed's model directory

Community & Resources

• One-take Creation, Flexible Referencing: Introducing Seedance 2.5 — the launch post, July 31, 2026, and the source of every 2.5 capability quote here
• ByteDance Seed model directory — canonical model list; as of August 2, 2026 Seedance 2.5 sits above 2.0
• Seedance 2.5 Ships: 30-Second Video in a Single Pass — our launch coverage
• Official Launch of Seedance 2.0 — the predecessor's launch post, February 12, 2026
• Seedance 2.0: Advancing Video Generation for World Complexity — the 2.0 technical report (arXiv 2604.14148, April 15, 2026): resolutions, durations, reference limits, Fast variant. No equivalent exists for 2.5
• arena.ai Image-to-Video leaderboard — dreamina-seedance-2.0-720p at #1; no 2.5 entry
• Artificial Analysis Text-to-Video leaderboard — 2.0 #1 with audio; no 2.5 entry
• ByteDance reportedly pauses global launch of Seedance 2.0 — TechCrunch, citing The Information
• MPA Sends Cease and Desist Letter to ByteDance Over Seedance 2.0 — The Hollywood Reporter
• Seedance 2.0: Breakthroughs and Copyright Launch Delay — our coverage of the dispute
• Compare with Sora 2, Kling 3.0, Wan 2.2, Hailuo 2.3, Vidu Q3 Pro, and Seedream 5.0