Seedance 2.5 API: Online video generation, tutorials and integration guide

What is Seedance 2.5

A native audio-video generation model from ByteDance

We position Seedance 2.5 as an AI audio-video generation model for video creation, advertising production, and multi-shot storytelling. ByteDance officially released this version on July 31, 2026, focusing on improvements to long-form video generation, multimodal references, and targeted editing of existing videos.

Unlike models that can only generate a simple animation from text, Seedance 2.5 aims to understand complete creative requirements. For example, you can ask it to generate a continuous sequence of a person walking from an office onto the street, while describing the person's actions, camera movement, ambient sounds, and transitions between different scenes.

We believe the main change in Seedance 2.5 is not just visual quality, but the opportunity for creators to complete more video production work in one go. Of course, it still cannot fully replace human direction and editing, especially when complex character interactions, strict dialogue, and precise movements are involved. The final visuals still need to be checked.

A storytelling model supporting 30-second generation and multiple rounds of video continuation

Seedance 2.5 increases the maximum duration of a single video generation from the previous generation's 15 seconds to 30 seconds, and supports extending stories through multiple rounds of continuation. This capability upgrade is relatively important for users who need to produce short dramas, narrative advertisements, or product promotional videos.

Previously, we might need to generate several 8-second or 15-second shots separately, then join them in editing software. This process frequently brings problems such as changes to characters' clothing, inconsistent lighting, or movements that do not connect between shots.

With Seedance 2.5, we can try describing the beginning, development, and ending of a story in a single 30-second clip. For example, first show a person entering a cafe, then have them walk to the counter and receive their coffee, and finally leave in a tracking shot.

We recommend not treating 30-second generation as a requirement every time. If you are only making a simple product close-up, 8 or 10 seconds may make the visuals easier to control and reduce generation costs. Videos lasting several minutes require multiple rounds of continuation; a single API request does not directly generate several minutes.

A video model supporting 50 multimodal reference assets and targeted editing

Seedance 2.5 can use up to 30 images, 10 videos, and 10 audio clips as references in a single generation, with no more than 50 assets in total.

We believe this capability is well suited to scenarios with clear visual requirements. For example, brand advertisements may need to reference product images, brand style, character appearance, scene layouts, and camera movement at the same time; short dramas may need to reference the appearances and voices of multiple characters as well as specific scenes.

You can use images to establish the appearance of characters and products, videos to describe the desired actions or camera work, and audio to provide sound references. The model combines these materials with the prompt to understand their relationships, rather than simply joining them together.

Seedance 2.5 also supports time-based content editing. You can ask the model to adjust actions, camera work, or backgrounds within a specific segment while preserving the other parts as much as possible. We believe this fits real production workflows better than regenerating the entire video for every change.

RequestBeam's Seedance 2.5 online tool and API

Through RequestBeam, we want to help you experience Seedance 2.5's real capabilities at a relatively low cost.

You only need to sign in to RequestBeam, create an API Key, and add a small amount of credit to submit video generation tasks, check their status, and preview results through the online tool. During evaluation, you do not need to develop a complete API integration first or purchase a fixed plan just to test a small number of videos.

For developers, we also provide a Seedance 2.5 API that lets you integrate video generation into your own products, automated workflows, or content production platforms. The current model identifier is seedance-2.5, accessed through a dedicated video generation endpoint.

We recommend completing a few real tests through the online tool first, determining whether the model meets your business needs, and then considering API integration. This can uncover issues with character consistency, generation duration, audio quality, and cost early, rather than discovering that the model does not suit the intended scenario after development is complete.

For practical use, RequestBeam can also reduce the hassle of maintaining accounts with multiple upstream model providers.

Seedance 2.5's core features and strengths

Long-form video storytelling: Telling a relatively complete story in 30 seconds

We believe one of the scenarios most worth trying with Seedance 2.5 is a short narrative video that requires continuous character performance.

One of our customers shared an experiment in which he used Seedance 2.5 to create a mother-daughter conversation lasting about 29 seconds. Rather than writing only a plot description, he prepared reference images of the mother, daughter, and setting, along with additional references for changes in the characters' emotions.

His prompt divided the content into character definitions, setting, dialogue, camera work, and a performance timeline. For example, a facial expression would only begin to change after a character delivered a particular line; as the argument escalated, the camera gradually moved closer and the characters' positions shifted as well. He was very satisfied with the result.

This is what we often describe to customers as a structured prompt. When making a 20-30 second narrative, we recommend first defining what needs to happen in each segment, then describing character actions, emotional changes, and camera movement. Compared with simply asking to "generate a cinematic emotional short film," this approach usually makes it easier to get good results.

Multimodal references: Keeping characters, products, and scenes consistent

Seedance 2.5's multimodal reference capabilities are well suited to content production that repeatedly uses the same characters or brand elements.

Another customer of ours used 5 audio references and added as many as 14 visual references in some video clips. The creator felt that the new version noticeably improved multi-character positioning, character performance, and continuity across shots compared with Seedance 2.0.

However, he also found that as the number of reference assets increased, the model began to struggle with complex relationships. In other words, supporting more assets does not necessarily mean the model can handle the relationships between them well.

For product advertisements, we recommend first providing front, side, and detail images of the product, then adding scene and camera references, while specifying which elements must not change, such as the logo, product color, packaging proportions, and button positions.

For character videos, we prioritize fixing the characters' appearance, clothing, and voices, then gradually add actions and scenes. Complex tasks can be tested in stages to avoid adding a large amount of material at once and then finding it difficult to identify which reference caused the result to deviate.

Native audio-video generation: Bringing dialogue, sound, and character performance together

Seedance 2.5 continues the direction of joint audio-video generation, allowing actions, dialogue, and ambient sound to be created in the same generation task.

We believe this capability suits character interviews, narrative short films, musical performances, and social media videos, because their effectiveness depends not only on attractive visuals but also on whether facial expressions, mouth movements, and voices are consistent when characters speak.

In accounts shared by customers, one user combined reference images of several characters with separate audio tracks to have 6 characters converse in the same scene. He was fairly satisfied with the overall performance, but also noted that some characters' voices changed and that laughter occurred earlier than expected, which felt somewhat strange.

We therefore recommend first testing voice assignment and lip synchronization with short videos before producing longer content in scenarios where character dialogue is particularly important. For advertising lines, brand presentations, or audio that must be accurate word for word, we also prepare a separate voice-over and audio post-production plan rather than relying entirely on a single generation.

Time-based editing: Making changes to specific shots

Seedance 2.5 supports describing shot changes using timing information, and also improves localized modifications, perspective adjustments, and green-screen replacement for existing videos.

For example, a 15-second product advertisement can be divided into three stages: the first 4 seconds show the complete product, the middle 7 seconds follow the product as it rotates, and the final 4 seconds push in toward product details. If the movement in the middle does not meet expectations, we can try requesting a change to that specific time segment rather than redesigning the entire piece.

We believe this capability is valuable for advertising and content production teams, because videos usually need multiple adjustments in real work. A client may want to retain a character's performance and only replace the background; or the product presentation may already work well, with only the camera movement needing adjustment.

It is important to note that time-based control does not mean frame-by-frame editing as in traditional editing software. The model may still change elements that were meant to remain unchanged, especially characters, product text, and background details in complex scenes.

We recommend changing only one clear target at a time and explicitly stating which content must be preserved. After the change, check not only the adjusted shot but also whether the actions and sound before and after it remain continuous.

Motion and visual texture: Suited to character close-ups and complex shots

Seedance 2.5 improves facial expressions, skin texture, lighting changes, and camera movement. We believe these details directly affect how natural a generated video looks.

One of our customers shared some details from a facial expression test with us. The character initially tried to suppress a smile, then gradually revealed amusement through changes in the lips, jaw, blinking, and gaze. The customer felt that the new version handled these subtle expressions better than he had expected.

We noticed that his prompt was very specific. Rather than simply asking the character to "look happy," he described the character first pressing their lips together, then changing their breathing, and gradually becoming unable to control their smile.

For short videos that need to convey a character's emotions, we recommend these descriptions of visible actions rather than only abstract adjectives.

Complex motion requires greater caution. ByteDance also acknowledges that the model still has room to improve stability in multi-character interactions and complex physical movements. For dance, fighting, sports movements, or shots requiring precise positional relationships, we therefore recommend adding motion references and paying particular attention to hands, bodily contact, and spatial relationships between characters. The more clearly you describe the details, the better the model's response tends to be. This is the conclusion we have drawn from feedback from multiple customers.

Seedance 2.5 benchmarks and performance

Here, we use Arena's public video model leaderboards alongside ByteDance's official specifications and the output parameters currently supported by RequestBeam for our analysis.

Text-to-video: No. 7 on Arena, scoring 1474±9

On Arena's Text-to-Video leaderboard, Seedance 2.5 720p ranks No. 7 among 48 models. Its score is 1474±9, with approximately 8001 votes in total.

Notably, Seedance 2.0 720p ranks No. 5 on the same leaderboard with a score of 1479±8. The difference is only 5 points, and the uncertainty intervals overlap, so this does not establish that Seedance 2.5 comprehensively surpasses the previous generation on ordinary text-to-video tasks.

We believe this result shows that Seedance 2.5's upgrades are not focused solely on basic visual quality. For common videos such as simple landscapes, rotating products, and single-character actions, Seedance 2.0 still deserves a place on the shortlist.

We recommend not switching simply because a new version is available if your current business only generates 5-10 second videos from a prompt. First compare the two generations using the same prompts and output specifications, then decide based on the proportion of final videos that are usable.

For work requiring longer narratives, multiple reference assets, or targeted editing, we would test Seedance 2.5 first, because these capabilities are more aligned with the direction of its upgrades.

Image-to-video: No. 4 on Arena, scoring 1477±7

On Arena's Image-to-Video leaderboard, Seedance 2.5 720p ranks No. 4 with a score of 1477±7 and approximately 15729 votes in total.

On the same leaderboard, Seedance 2.0 scores 1475±7 and ranks No. 5, while MiniMax H3 ranks No. 1 with 1495±5.

Based on these results, we believe Seedance 2.5 is already a competitive image-to-video model, but its difference from Seedance 2.0 is not pronounced.

In practical business use, the value of image-to-video often depends on whether the model preserves important information in the input image. For example, does a product's shape stay consistent, does a person's face change, and does text or a logo in the image become distorted?

We recommend treating product presentations, animated character portraits, and animated illustrations as separate test scenarios. For ordinary image animation, you can compare Seedance 2.5 with MiniMax H3; for videos that require multiple images to jointly control characters and scenes, focus on evaluating Seedance 2.5's reference capabilities.

Video editing: No. 2 on Arena, scoring 1410±26

On Arena's Video Edit leaderboard, Seedance 2.5 ranks No. 2 with a score of 1410±26. Wan 3.0, ranked No. 1, scores 1414±26, while Seedance 2.0 ranks No. 5 with 1365±14.

These results suggest that Seedance 2.5's improvement over the previous generation is more noteworthy for video editing tasks.

However, this result needs to be interpreted cautiously. Seedance 2.5 has only 429 votes on this leaderboard, and the uncertainty in its score is relatively large, so it cannot yet be taken to mean that it consistently outperforms other models on all video editing tasks.

We recommend prioritizing tests with Seedance 2.5 if your business involves replacing video backgrounds, modifying localized character actions, redesigning camera movement, or creating variations from reference videos.

During evaluation, we recommend focusing on two things: first, whether the model correctly completes the specified change; second, whether content that was not supposed to change remains stable. For commercial advertisements, the second condition may matter more than how attractive the generated visuals are.

Long-form video performance: Single-generation duration increases to 30 seconds

ByteDance officially confirms that Seedance 2.5's maximum duration for a single generation is 30 seconds, compared with 15 seconds for Seedance 2.0. The new version also supports multiple rounds of continuation to connect multiple video clips into a longer story.

We believe this is a specification upgrade with clear business value, but it does not mean the model can maintain the same stability in every frame across 30 seconds.

For narrative videos, longer generation durations increase the likelihood of changes in character features, unnatural action transitions, or deviations in audio pacing. Continuous dialogue and scene transitions involving multiple characters are more difficult than a scene with a single product slowly rotating.

We recommend testing 10-second, 15-second, and 30-second versions of the same story, comparing character consistency, completeness of actions, shot transitions, and the parts that ultimately need manual changes.

If a 30-second video can be used directly, it can indeed reduce stitching work in post-production. If longer clips have a noticeably higher failure rate, we instead recommend shortening the generation duration.

Output specifications and generation costs: Up to 1080p, with costs evaluated by video duration

RequestBeam's current Seedance 2.5 endpoint supports 480p, 720p, and 1080p output. Video duration can be set to 4-30 seconds, and an automatic duration option is also available.

We believe that comparing only the per-second price is not enough for video models. If a video needs to be generated four times before the result is satisfactory, the actual production cost must account for those attempts, not just the final successful video.

We recommend prioritizing 480p or 720p during early testing, confirming that the prompt, characters, and camera work meet your requirements, then evaluating whether 1080p is necessary. Note that a low-resolution preview and a high-resolution generation do not necessarily produce exactly the same visuals.

For batch e-commerce videos, short-video automation, and advertising asset production, we recommend recording the total cost of each task, the number of generations, the proportion of usable videos, and the time spent on post-production changes before deciding whether to expand usage.

Comparing Seedance 2.5 with Seedance 2.0, Veo 3.1, Kling 3.0, and MiniMax H3

Compared with Seedance 2.0: A more worthwhile upgrade for long narratives and targeted editing

The most direct differences between Seedance 2.0 and Seedance 2.5 are single-generation duration, multimodal reference capacity, and editing capabilities.

Seedance 2.0 already supports reference-based generation using images, video, and audio, while Seedance 2.5 further increases the single-generation duration to 30 seconds and expands reference capacity to a maximum of 50 assets, while strengthening time-based editing and complex camera control.

However, Arena's text-to-video and image-to-video leaderboards show very similar basic generation results for the two generations. For text-to-video, 2.0 even scores slightly higher than 2.5.

We do not recommend treating Seedance 2.5 as a version that every task should automatically upgrade to.

If you currently use Seedance 2.0 to generate short videos, product animations, or simple character actions and the results are already stable, we recommend retaining it as an option for low-complexity tasks.

If your main problems are insufficient video duration, difficulty connecting multiple shots, or needing to regenerate an entire video just to change one local element, Seedance 2.5 is more worth trying.

Compared with Google Veo 3.1: Choosing between complex references and high-resolution workflows

Google Veo 3.1 also supports native audio-video generation, image references, video extension, and first- and last-frame control.

According to Google's public API documentation, Veo 3.1 can generate 4-, 6-, or 8-second videos and supports up to 3 reference images. Google also offers 1080p and 4K output options, with high resolutions using enhancement techniques and subject to specific duration and generation-mode restrictions.

We believe Veo 3.1 is more worth using in scenarios with clear requirements for image resolution, Google's developer ecosystem, and short-shot delivery. For example, a brand may want to produce high-resolution presentation videos, or may already use the Gemini API and Vertex AI for content production.

Seedance 2.5 is better suited to putting a longer story, multiple character references, and multiple shot requirements into a single generation task. It can generate up to 30 seconds at once and supports more types of reference assets.

We recommend testing both Veo 3.1 and Seedance 2.5 if you mainly produce high-resolution commercial shots of 8 seconds or less; if you need a continuous 20-30 second narrative, prioritize evaluating Seedance 2.5.

Compared with Kling 3.0: Different strengths in long-shot storytelling and high-resolution commercial production

Kling 3.0 is a video model series released by Kuaishou in 2026, supporting multi-shot storytelling, character consistency control, and native audio-video generation. Its officially announced maximum single-generation duration is 15 seconds, with enhanced multilingual dialogue capabilities.

The Kling 3.0 series subsequently added native 4K video generation. We believe this gives it an advantage worth noting in scenarios requiring high-resolution commercial delivery.

For Seedance 2.5, we would focus more on its 30-second single generation, multimodal references, and editing of specific time segments. For Kling 3.0, testing can focus on high-resolution visuals, character dialogue, and performance under explicit storyboard requirements.

For example, when making a luxury product advertisement with only a few high-quality shots, you can prioritize comparing Kling 3.0's high-resolution generation mode. If you need a 25-second video in which a person moves from indoors to outdoors while continuously conveying a complete event, Seedance 2.5's duration limit is more convenient.

We recommend not judging the visual quality of the two models only from release samples. Instead, use the same character images, story requirements, and acceptance criteria to compare whether final delivery requires extensive manual repairs and rework.

Compared with MiniMax H3: Image-to-video and brand text rendering deserve particular comparison

MiniMax H3 was released on July 31, 2026. It supports multimodal inputs including text, images, video, and audio, and can generate audio-video content up to 15 seconds long at resolutions up to 2K. Its official introduction highlights brand text rendering, instruction following, and motion transfer.

On Arena's image-to-video leaderboard, MiniMax H3 ranks No. 1 with 1495±5, while Seedance 2.5 ranks No. 4 with 1477±7. This shows that MiniMax H3 has some advantage in this independent preference test.

We believe MiniMax H3 deserves to be tested alongside Seedance 2.5 if your business mainly turns a single product image into a short video, or if you have demanding requirements for brand text, product packaging, and visual detail.

On the other hand, Seedance 2.5 supports longer videos in a single generation and a larger number of reference assets. For projects involving multiple characters, multiple settings, and complex relationships between shots, these features may reduce asset organization and video stitching work.

We recommend assigning the two models by scenario: prioritize comparing MiniMax H3 for short product animations and brand assets, and Seedance 2.5 for longer narratives, multiple reference assets, and localized editing. If your business includes both types of tasks, retaining two models may make more sense than using one model for everything.

Meanwhile, MiniMax H3's low cost is also a strong competitive advantage.

Seedance 2.5 API FAQ

How do I call the Seedance 2.5 API through RequestBeam?

We recommend first signing in to RequestBeam and creating an API Key, then reading our documentation or asking AI to help you configure the connection.

We recommend completing a minimal generation test through the online tool first, then integrating according to RequestBeam's current Seedance 2.5 API documentation. This lets you verify whether the model meets your business needs before integration and helps avoid mistakenly carrying over endpoint paths and parameters from other providers into RequestBeam.

Why does the Seedance 2.5 API return parameter errors?

We recommend first checking the generation duration, resolution, aspect ratio, and whether you are using parameters unsupported by the current endpoint.

In RequestBeam's current Seedance 2.5 endpoint, duration can be an integer from 4-30, or you can use -1 to let supported workflows select the duration automatically. Automatic mode does not guarantee that every generation will be 30 seconds long.

resolution only accepts 480p, 720p, or 1080p. The current endpoint does not support 4k; even if other models or video platforms offer similar options, they cannot be reused directly here.

aspectRatio supports 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, and adaptive. If you use images with the first_frame or last_frame role, you need to use the adaptive aspect ratio.

We recommend not copying every parameter directly when migrating older Seedance requests. Validate a minimal request first, then add configuration options one at a time. Also avoid submitting duplicate parameter aliases that mean the same thing but have inconsistent values, reducing unnecessary 400 errors.

How do I upload image, video, and audio references, and why do audio references cause errors?

RequestBeam's Seedance 2.5 endpoint supports providing reference assets through imageUrls, videoUrls, and audioUrls. You can also submit information with types and roles through metadata.content.

The currently supported limits are 30 images, 10 videos, and 10 audio clips, with no more than 50 assets in total. Asset addresses can be accessible HTTP(S) URLs or existing asset:// identifiers.

When reference content is relatively complex, we recommend clearly distinguishing its purposes, such as character appearance, setting, actions, and sound. For example, you can specify first- and last-frame roles using imageWithRoles, while ordinary reference images should not mistakenly be used as first or last frames.

It is particularly important to note that the current endpoint requires audio references to be submitted together with at least one image or video. If there are only audioUrls without corresponding visual assets, the request may be rejected.

If you need the model to generate sound, you can use the metadata.generate_audio boolean parameter. We recommend first validating three modes separately: no audio, generated audio, and audio references, before evaluating complex multi-character dialogue scenarios.

Why is there no video after a successful API response, and why can't an existing video use draft_task?

Seedance 2.5 uses asynchronous tasks. A successful submission only means that the service has accepted the task, not that video generation has finished.

RequestBeam returns a task identifier, jobId, and status information. We recommend saving this identifier and using GET https://api.requestbeam.com/v1/jobs/{jobId} to check generation status.

Once the task status is COMPLETED, read the video result from data.outputResult. Your parser needs to accommodate video addresses returned as either a string or an array, rather than assuming that the returned data always has only one structure.

In addition, Seedance 2.5 supports reusing eligible existing drafts through draft_task, but not every completed video can be used directly as a draft.

According to the current documentation, the draft must come from the same account, use the same model and processing channel, have a completed status, and have been created less than 7 days ago. Reuse requires the corresponding task identifier, and new prompts, durations, reference assets, and similar parameters cannot be submitted at the same time. Output resolution is among the settings that can be adjusted.

We recommend designing ordinary video generation and draft reuse as two clearly separate API flows, with dedicated handling for task timeouts, failures, incompatible assets, and result address parsing. For products that need to generate videos in batches, these engineering details directly affect the actual success rate and operating costs.