Seedance 2.5

Scroll sideways for more →

Seedance 2.5 is a proprietary AI video generation model released by ByteDance on July 31, 2026. It jointly generates audio and video from combinations of text, image, video, and audio inputs. ByteDance positioned it as the next named release in the Seedance family after Seedance 2.0, with longer one-pass output, a larger reference budget, multi-round extension, and more targeted editing.^[1][2]

The model can produce up to 30 seconds of audio-video content in one generation. A request can include as many as 30 images, 10 video clips, and 10 audio clips, for a maximum of 50 reference assets. ByteDance also documents timestamp-directed changes, green-screen editing, camera-perspective changes, and the use of clay or untextured 3D renders to guide staging and camera movement.^[1]

Seedance 2.5 was released as a hosted model, not as downloadable weights. It first rolled out through ByteDance consumer products, including Jimeng AI and Doubao Pro, followed by international API access through BytePlus ModelArk. ByteDance had not published a Seedance 2.5 technical report or announced a weights release by August 19, 2026. The available technical evidence therefore describes the model's behavior, serving interface, and training-data categories, but not its parameter count, training compute, detailed architecture, or inference requirements.^[1][2][5][6]

Release and access

ByteDance Seed's July 31 launch article called Seedance 2.5 a new-generation video creation model. The release began rolling out through Jimeng AI, the Chinese product related to Dreamina, and the professional tier of Doubao. Caixin independently reported the same release date and product rollout. At launch, ByteDance said an API would follow through its ModelArk service.^[1][14]

BytePlus announced availability through ModelArk on August 6. It used the product name Dreamina Seedance 2.5 for the hosted international offering. BytePlus said access was available in supported markets and explicitly noted that BytePlus was not available in the United States. This is a regional hosted service; it is not a license to download, modify, or self-host the model.^[6]

The official Seedance project page and the BytePlus activity page also provide browser-based trial paths. These product surfaces may expose different controls or output tiers. A feature shown in Jimeng, Doubao, or a BytePlus playground should not automatically be assumed to exist in every API region or third-party integration.^[2][7]

Generation and editing capabilities

The main change from Seedance 2.0 is the length of a single generation. ByteDance says Seedance 2.5 can create a 30-second sequence with synchronized audio in one pass. It is designed to organize multiple connected shots inside that interval rather than merely prolonging one unchanged scene. The launch examples show prompts divided into narrative stages and camera moves, but those examples were selected by the developer and are demonstrations rather than a controlled success-rate study.^[1]

The model also supports repeated extension. A generated sequence can be supplied as a reference for a later generation, with instructions to continue its characters, setting, pacing, and sound. ByteDance describes this as multi-round extension and says it can be used to build content longer than one 30-second output. The company did not publish a tested upper bound, continuity failure rate, or independent comparison for multi-round use, so the capability is better understood as an iterative workflow than as a guarantee of stable multi-minute generation.^[1]

Reference capacity increased across all three media types. ByteDance lists separate maxima of 30 images, 10 video clips, and 10 audio clips in one request. References can supply character appearance, voice, composition, setting, props, motion, camera language, or sound. The model is a multimodal model, but the presence of 50 inputs does not mean all inputs receive equal influence. ByteDance has not published an ablation study measuring how performance changes as the number or mixture of references increases.^[1]

Editing is integrated into the same model family. During generation, a prompt can assign events or camera instructions to particular time ranges. After generation, users can request changes to a character, action, background, lighting, product, or camera treatment while asking the system to retain the rest of the sequence. The launch materials also describe green-screen replacement, reference-based editing, and control from a clay or white 3D model. In that workflow, the render supplies spatial layout, poses, motion paths, and camera angles, while other references supply appearance and lighting.^[1][2]

These are developer-described capabilities, not published reliability measurements. ByteDance does not report how often an edit changes supposedly protected regions, how reference conflicts are resolved, or how precisely requested timestamps are followed across a representative prompt set.^[1]

Relationship to Seedance 2.0

ByteDance says Seedance 2.5 builds on the unified multimodal audio-video joint-generation architecture introduced with Seedance 2.0. The 2.0 technical report describes that earlier system in detail, but no corresponding report for 2.5 was public at the evidence cutoff. It is therefore accurate to describe architectural continuity, but not to assume that every 2.0 component, training recipe, or performance result remained unchanged.^[1][5]

Scroll sideways for more →

The comparison uses the open-platform limits in the Seedance 2.0 report and the maxima in the 2.5 launch article. It shows a doubled maximum one-pass duration and an increase from 15 to 50 when the three modality-specific reference maxima are added. It does not establish that 2.5 is better on every task. For example, an increased input allowance says nothing by itself about output fidelity, physical realism, latency, or cost.^[1][5]

Resolution and serving boundaries

The launch and serving materials were inconsistent about resolution at the August 19 cutoff. BytePlus's activity-page FAQ and independent July 31 reporting described 480p and 720p tiers. Arena likewise evaluated a model identified specifically as dreamina-seedance-2.5-720p.^{[7][10][11][12][15]}

On August 14, BytePlus separately announced native 1080p output with native 10-bit color. When the BytePlus plan page was checked on August 19, its plan cards listed 480p, 720p, and 1080p, while the FAQ on the same page still said that 720p was the maximum. The most cautious interpretation is that BytePlus had announced a 1080p rollout while its documentation and access surfaces were still being updated. The announcement does not demonstrate that 1080p was enabled in every region, product, or API mode.^[7][8]

No checked first-party Seedance 2.5 source documented 4K output. Third-party pages and social posts making 4K claims are not a basis for a general model specification. Similarly, the hosted model identifier, prices, and plan allowances may change independently of the underlying model release.^[2][7]

Training-data disclosure

ByteDance published a six-page training-data summary for Seedance 2.5 under the European Union's general-purpose AI transparency format. The document is version 1.0, dated July 31, 2026, and identifies ByteDance Nexus AI Pte. Ltd as the provider. It reports that the model was placed on the European Union market in July 2026 and that the latest training-data acquisition or collection occurred in June 2026.^[3][4]

The size figures are broad bands rather than exact counts. ByteDance reports between 1 billion and 10 trillion text tokens, more than 1 billion images, more than 1 million hours of audio, and more than 1 million hours of video. It says most audio came from associated video files and only a limited subset was standalone audio. The document does not identify model parameters, training steps, compute, data deduplication rates, or the exact share contributed by each source class.^[4]

According to the provider, the training mixture combined publicly available data, licensed data, and synthetic data. The summary says public image and video datasets were used but does not name the large datasets. It also reports commercial licensing agreements covering image, video, and audio, along with private third-party datasets obtained on a licensed basis. These statements disclose source categories, not the identities of all rightsholders or datasets.^[4]

ByteDance says its Bytespider crawler collected publicly accessible images, video, and audio for training, testing, and validation through June 2026. The company states that the crawler is designed to respect robots.txt, avoid paywalls and password-protected material, and avoid overloading sites. The same summary says neither user interactions with Seedance 2.5 nor interactions with ByteDance's other products or services were used to train the model.^[4]

Synthetic data consisted of text descriptions of images and videos. ByteDance says these descriptions were generated with vision-language models, large language models, and potentially fine-tuned internal models. The company also reports preprocessing and filtering intended to remove unsafe or harmful material. It states that it was not a signatory to the general-purpose AI Code of Practice section covering reservations of rights under text-and-data-mining exceptions. The transparency summary is a company disclosure and was not presented as an external audit.^[4]

Arena evaluation

The strongest independent quantitative evidence available shortly after release came from LMArena, which now operates as Arena. Its video leaderboards use pairwise human preferences and a Bradley-Terry ranking model. Arena publishes score intervals and a rank spread intended to show which ranks remain plausible when confidence intervals overlap.^[9][13]

Arena's pages were dated August 14, 2026 and evaluated the 720p hosted variant. They reported the following snapshot:

Scroll sideways for more →

The intervals are central to interpreting the raw ranks. Seedance 2.5 had the highest video-edit point estimate, but its 1-2 rank spread overlapped the next model. Its image-to-video spread included ranks 1 through 3, and its text-to-video spread included ranks 2 through 6. Arena's open ranking example computes 95 percent confidence intervals at a 0.05 significance level. The snapshot therefore supports strong early user preference, especially for editing and image-conditioned generation, but not a statistically unambiguous claim that the model was the sole best system in those categories.^{[9][10][11][12][13]}

The evidence was also young. Seedance 2.5 had 403 video-edit votes, 2,912 image-to-video votes, and 1,337 text-to-video votes in its rows, fewer than several older systems. Human-preference ratings can move as the sample grows. They do not directly measure latency, price, prompt adherence, physical correctness, audio fidelity, rights compliance, or how often an edit preserves unrequested regions.^[10][11][12]

Safeguards and rights issues

BytePlus says its hosted Seedance 2.5 service applies content filters, visible watermarks, and C2PA Content Credentials. It also describes controls intended to block unauthorized copyrighted-character generation and to restrict generation from real-person images or video. For vetted clients, the service offers a process for identity verification and likeness authorization. These are descriptions of BytePlus's service layer; they do not establish that every third-party deployment has the same controls.^[6]

The release followed public disputes about copyright and likeness controls in Seedance 2.0. On August 17, Reuters reported that ByteDance and the Motion Picture Association signed an agreement to strengthen safeguards for Seedance and Seedream. The agreement followed complaints from film studios that the systems could generate copyrighted characters and celebrity likenesses without authorization. ByteDance said newer model versions had stronger protections, and both parties said they would continue working on safeguards.^[16]

The agreement and product filters do not answer every question about training data, nor do they guarantee that infringing or deceptive output is impossible. The training-data summary discloses broad source categories, while the BytePlus controls govern use of the hosted service. Those are related but distinct layers of evidence.^[4][6][16]

Limitations and open questions

ByteDance itself identifies two remaining weaknesses: the physical plausibility of complex motion and stability when multiple subjects interact. These limits are important because long, multi-shot sequences create more opportunities for identity drift, inconsistent geometry, and physically implausible transitions. The official examples do not report failure rates or show randomly sampled outputs.^[1]

Other important unknowns remained at the cutoff. ByteDance had not published a 2.5 model card with architecture, parameter count, compute, energy use, detailed evaluation prompts, or per-capability success rates. There was no public weights release, and no independent academic paper or replication specific to Seedance 2.5 had been identified. The early Arena results cover anonymous preference for a 720p hosted variant, not the full set of product modes.^{[1][2][4][10][11][12]}

Seedance 2.5 is therefore best described as a distinct, commercially deployed extension of the Seedance family whose longer output, larger reference budget, and editing interface are well documented. Claims about its internal design, universal resolution ceiling, self-hosting, or superiority outside the dated preference snapshots remain unsupported.^[1][2][7]

References

1. ^ByteDance Seed, "One-take Creation, Flexible Referencing: Introducing Seedance 2.5," July 31, 2026. seed.bytedance.com/...ing-introducing-seedance-2-5
2. ^ByteDance Seed, "Seedance 2.5" model page. Accessed August 19, 2026. seed.bytedance.com/...seedance2_5
3. ^ByteDance Seed, "Transparency." Accessed August 19, 2026. seed.bytedance.com/transparency
4. ^ByteDance Nexus AI Pte. Ltd, "Public Summary of Training Data Content for Seedance 2.5," version 1.0, last updated July 31, 2026. lf3-static.bytednsdoc.com/...%20Seedance%202.5.pdf
5. ^Team Seedance et al., "Seedance 2.0: Advancing Video Generation for World Complexity," arXiv:2604.14148, 2026. arxiv.org/...2604.14148
6. ^BytePlus, "Dreamina Seedance 2.5 is Now Available on BytePlus: Advancing Controllable, Enterprise-Ready Video Creation," August 6, 2026. byteplus.com/...dreamina-seedance2-5
7. ^BytePlus, "Dreamina Seedance 2.5" activity and plan page. Accessed August 19, 2026. ai.byteplus.com/...seedance2-5
8. ^BytePlus Global, "Native 1080P is here on Dreamina Seedance 2.5," X post, August 14, 2026. x.com/...2088156784250986884
9. ^Arena, "Arena's Ranking Method," updated February 28, 2026. arena.ai/...ranking-method
10. ^Arena, "Video Edit Arena," leaderboard snapshot dated August 14, 2026. arena.ai/...video-edit
11. ^Arena, "Image-to-Video Arena," leaderboard snapshot dated August 14, 2026. arena.ai/...image-to-video
12. ^Arena, "Text-to-Video Arena," leaderboard snapshot dated August 14, 2026. arena.ai/...text-to-video
13. ^Arena, "Arena-Rank: Open Sourcing the Leaderboard Methodology," updated February 21, 2026. arena.ai/...arena-rank
14. ^Guan Cong, "ByteDance and MiniMax Roll Out Upgraded AI-Video Models," Caixin, July 31, 2026. companies.caixin.com/...102469970
15. ^Dingjiao One, "MiniMax and ByteDance Make Same-Day Moves as Video Models Enter a New Stage," Jiemian News, republished by Sina Finance, July 31, 2026. finance.sina.com.cn/...doc-iniktkau9209815.shtml
16. ^Rashika Singh, "ByteDance Signs AI Copyright Pact with Hollywood Trade Group," Reuters, August 17, 2026. wincountry.com/...-pact-with-hollywood-trade-group