ByteDance (Seedance)

Seedance is the video side of ByteDance Seed, ByteDance's internal AI research division. ByteDance, as the parent of TikTok, Douyin, and CapCut, sits on the world's largest dataset of short-form video with paired engagement data: view counts, completion rates, likes, shares, comments, and replay patterns across billions of clips. They know not just what videos look like but also which ones people actually watch, which gives them a structural training-data advantage that no other lab has been able to replicate. The fal/a16z report covered in the industry-data chapter explicitly flagged Seedance 1.0 as topping the Artificial Analysis video leaderboards in June 2025 and noted that 'the previews of Seedance 2.0 are blowing us away,' which is an unusually strong endorsement from investors who talk to every video model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google..

Unlike Kuaishou (which built Kling as a consumer product first and then broadened it), ByteDance's video work serves two parallel customers. The first is ByteDance itself: TikTok, Douyin, and CapCut users generate millions of AI-enhanced clips per day, and every one of those generations runs on ByteDance's internal Seedance models. The TikTok use case requires sub-second generation at scale, which forces Seedance to be unusually efficient at inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. time. The second customer is external developers who pay to use Seedance through ByteDance's APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. and through partner platforms like Krea, fal.ai, and Replicate. The external business is smaller than the internal usage but growing, and it is the reason ByteDance has been increasingly willing to publish technical details, participate in public benchmarks, and allow Western operators to access the models at all. The Seedance research team does not do much English-language press, and individual researcher names are much less known in the Western AI community than the work itself would suggest they should be.

Seedance uses what ByteDance calls a 'decoupled spatiotemporal transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks.' architecture built around a 'time-causal VAEVAE (Variational Autoencoder)A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable..' These terms are worth unpacking because they describe real design decisions. The time-causal VAE compresses video into a 3D latent spaceLatent spaceThe compressed numeric representation an image is turned into before diffusion. Working here instead of on raw pixels is what makes modern image models fast enough to run. where each temporal frame can only depend on frames that came before it, not on future frames. This is the opposite of how most 3D VAEs work (which use bidirectional attention across time) and it is specifically useful for streaming applications where you want to generate video frame-by-frame as it plays rather than producing the entire clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. up front. The decoupled spatiotemporal transformer means the model has separate but coupled attention paths for spatial structure (what is in each frame) and temporal structure (how frames relate to each other over time), rather than treating time as just another dimension like Sora does. The trade-off is that Seedance gives up some raw coherence in exchange for dramatically better efficiency and more flexible generation modes.

The other major architectural investment in Seedance is aggressive distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.). and what ByteDance calls video-specific reinforcement learning from human feedback. The distillation work has pushed Seedance 1.0 to generate 5-second clips in about 41 seconds of compute time on production infrastructure, which is roughly 5 to 10 times faster than Veo 3 or Sora 2 for comparable output quality. The RLHFRLHF (Reinforcement Learning from Human Feedback)Training that tunes a model toward outputs humans prefer, using human ratings as the reward signal. loop is trained on preference data from ByteDance's consumer products, where users can thumbs-up or thumbs-down the AI-generated content they see in TikTok and CapCut. Those preference signals get fed back into the model as a reward signal on top of the diffusion training, which is the same pattern Midjourney uses for images. The effect is that Seedance's aesthetic preferences reflect the accumulated taste of TikTok's global user base, slightly more cinematic, more dynamic, and more visually immediate than the average model trained on scraped web data.

Seedance 2.0, released in early 2026, introduced what ByteDance calls a 'unified multimodal audio-video joint generation architecture' that supports text, image, audio, and video inputs in a single model. The headline capability is comprehensive multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. reference and editing, meaning you can give Seedance 2.0 multiple kinds of inputs at once (a starting image, a reference video for camera motion, an audio file for timing, a text promptPromptThe text description you provide to a model to specify what you want it to generate. for scene description) and it will generate output that respects all of them. This is a significant architectural step beyond the single-modality input approach most video models use, and it is the primary reason Seedance is now the character-consistency leader, you can feed it a character design image and have it show up reliably in every generation.

Seedance 1.0 (June 2025)

The original release. 1080p output at 24 fps, multi-shot narrative support where the model can generate sequences of related shots with consistent characters and scenes, and notably aggressive distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.). that brings 5-second video generation down to about 41 seconds of compute time. Topped the Artificial Analysis video benchmarkBenchmarkA standardized test used to compare models on a task, such as prompt adherence or image quality. on release, surpassing Hailuo and Veo 2. This was the first widely-praised feature of Seedance: both fast and high quality, which is unusual.

Seedance 1.5 Pro (late 2025)

A joint audio-video model with improved instruction following and motion quality. This was the version where ByteDance caught up to Veo 3 and Sora 2 on the audio-video front, delivering synchronized dialogue, sound effects, and ambient audio alongside the visual content.

Seedance 2.0 (early 2026)

The current flagship. The unified multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. architecture is the headline feature, but the version-over-version quality improvements are just as significant: better motion coherence over longer clips, better character persistence across multi-shot sequences, improved prompt adherencePrompt adherenceHow faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood., and expanded resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. support. Seedance 2.0 is available through ByteDance's APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. and through partner platforms including fal.ai (as of April 9, 2026), Krea, and Replicate. In the current operator playbook, Seedance 2.0 is the default choice for any scene that requires a specific character or product to appear consistently across multiple generations. The global rollout has been complicated by a significant copyright controversy: Seedance 2.0 went viral for clips featuring recognizable actors without permission, prompting the Motion Picture Association to denounce ByteDance for copyright infringement, cease-and-desist letters from The Walt Disney Company, accusations from Paramount of blatant infringement, and a letter from US Senators Blackburn and Welch asking ByteDance's CEO to shut down Seedance. Despite this, over 600 people attended the official US premiere event in Hollywood on April 17, 2026, which included the first-ever Golden Seed AI Film Awards. The copyright situation is unresolved and represents a real risk for operators building on Seedance in Western markets.

Seedance's strengths are character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work., unified multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. input handling, efficient inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training., and the aesthetic advantage that comes from training on TikTok's engagement data. For any work that requires the same character to appear in multiple shots with consistent identity, Seedance, now at version 2.5, is the best single choice for this as of mid-2026. For any work that needs to condition on multiple reference inputs at once (image plus style reference plus audio timing), Seedance is the only frontier model that handles all of those natively in a single generation. The cost per generation is also lower than Veo 3.1 or Kling 3.0 at comparable quality, which matters for high-volume production use cases. And the engagement-trained aesthetics produce video that feels visually immediate and well-composed for vertical feeds, which is exactly what a lot of commercial production work needs.

Seedance's weaknesses are geographic availability, tooling maturity in Western markets, and the question of long-term strategic commitment. ByteDance's primary customer is ByteDance itself, and the external APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. business exists partly as a way to build research credibility and partly as a way to monetize spare capacity. If ByteDance ever decides the external business is not worth the operational overhead, the API could be discontinued or made China-region-only, similar to what happened with some earlier ByteDance research releases. The documentation is also less polished than Western alternatives, and first-class tooling support in ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. and the major inference platforms exists but lags behind Kling and Veo. For operators building pipelines in Western markets, these are real concerns that need to be weighed against the capability advantages.

ByteDance's strategic position in video generation is the most structurally advantaged in the entire field, and everyone in the industry knows it. TikTok provides training data that no other lab can match, the consumer products provide distribution and inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. volume that amortizes compute costs, and the parent company has the capital and patience to fund frontier research for as long as it takes. The result is that Seedance will probably remain near the top of video model quality leaderboards for the foreseeable future, because there is no plausible scenario where ByteDance loses the data advantage and no plausible scenario where they lack the compute to ship frontier models. The only question is how willing ByteDance is to make those models accessible to Western operators versus keeping them captive to the TikTok and CapCut ecosystem.

For operators, Seedance is the clearest case of a frontier video model that you should use when the quality and the multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. input handling justify accepting the geopolitical and operational uncertainty. Route hero character-consistent work to Seedance 2.0, keep a fallback to Wan 2.6 or Kling 3.0 in case the access model changes, and budget for the possibility that your Seedance pipeline may need to migrate on short notice if ByteDance changes its external APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. strategy. This is not a reason to avoid Seedance, it is a reason to have a migration plan in place while you use it.

The confusing thing about using Seedance is that you often are not told you are using Seedance. The same underlying model reaches people under at least four names: Jimeng in China, Dreamina internationally, the Doubao assistant, and, since March 2026, inside CapCut, where Dreamina Seedance 2.0 was folded into the editor that hundreds of millions of people already have on their phones. For developers the path is BytePlus ModelArk and Volcengine, which is also where the real caveat lives: ByteDance's video service terms have stated it is not available in the United States, so US builders often reach Seedance through third-party hosts like fal or Replicate instead of directly.

On the craft, Seedance rewards structure. Its genuine differentiator is native multi-shot generation, so prompts written as explicit shot-by-shot beats get the most out of it, and naming the camera move, the lens, and the lighting per shot pays off the way it does with a real crew. For image-to-videoImage-to-videoStarting from a still image and generating a video that animates it according to a prompt., a strong first frame conditions the motion heavily, so spend effort on that frame. It is at its best on promptPromptThe text description you provide to a model to specify what you want it to generate.-faithful cinematic short clips and multi-shot mini-narratives with consistent characters, and at its weakest on long single takes and any scene that needs reliable, legible on-screen text.

The direction is unified multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. generation: one model that takes text, image, audio, and video as inputs and returns video with synchronized audio, which is exactly the jump Seedance 2.0 made when it added native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward. that the original Seedance 1.0 report did not have. Expect longer and higher-resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. clips, tighter integration into CapCut and Doubao, and continued aggressive cost reduction through distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.)., because the whole ByteDance strategy depends on each generation being cheap enough to give away inside consumer apps. Reports of a Seedance 2.5 with longer, native-4K output and a Seedream 5 image model exist, but as of this writing no ByteDance primary source or tier-1 outlet confirms them, so treat the specific next releases as rumor rather than roadmap.