Understanding Bytedance’s Video Generators
Last week, a new product appeared on the video scene — Seedance 1.0 video generator. It looks impressive and even surpasses Veo3 in terms of specifications. However, it’s not yet possible to test it in full capacity. Good news: an accessible version, Seedance 1.0 Mini (also known as Video 3.0), is available on the website https://dreamina.capcut.com/ai-tool/home, which allows users to generate up to 120 videos per day for free using credits.
Additionally, it was announced yesterday that Fal.ai permits trial access to a reduced version of SeeDance 1.0. The release date and availability of the full, “adult” version are still unknown. All demonstration videos circulating on social media are taken from platforms where users test new features randomly without full understanding.
Last Friday, I discussed Seaweed APT2 — the most powerful (but still unavailable) video generator from the same company. It can create real-time videos at 736×416 resolution on a single H100 GPU and 1280×720 on four H100 GPUs, with durations up to one minute.
In January, a simpler version was already available — Seaweed APT1, which could generate sessions up to 2 seconds in real time.
To clarify: Seaweed APT2 is an advanced version of the SeeWeed model (discussed as early as March 2025). The term APT stands for Autoregressive Adversarial Post-Training — a training method based on autoregressive and adversarial approaches. This model, with 8 billion parameters, can stream video in real time at 24 frames per second, without stopping, with a resolution of 736×416, generating clips in a single step. Details, such as the use of KV-cache for long generation, are available at: https://seaweed-apt.com/2 and https://seaweed-apt.com/1.
Regarding Seedance: it is a distilled version of the SeeWeed model, created using the APT method — meaning it is simpler, lighter, and more optimized than the original. Essentially, the main family is SeeWeed, with Seedance being its enhanced and streamlined version for video generation.
Additionally, they have the SeedVR2 video upscaler, based on diffusion and adversarial methods, designed for restoring and improving video quality in one step. I wrote about this last week.
Furthermore, the company offers SeedEdit 3.0 — an advanced AI-powered image editing tool that preserves details well. It allows prompt-based modifications, and compared to previous versions 1.0 and 1.6, released earlier, version 3.0, launched on June 5, has become much more powerful.
ByteDance also has its own OCR model, Dolphin, designed for recognizing text in PDF documents — capable of identifying all static objects. Repository: https://github.com/bytedance/Dolphin, demo version: https://huggingface.co/spaces/ByteDance/Dolphin.
Finally, they offer CameraCtrl II — a tool for managing cameras and video streams.
If all of this seems too complex, I recommend just visiting https://seaweed-apt.com/ — everything there is designed to be extremely user-friendly.
In my view, Bytedance is like a Chinese Google Labs, with numerous experimental projects and ideas. Keep an eye on them — a full, mature Seedance 1.0, which will be widely adopted, is coming soon.