ByteDance / Seedance 2.0 Fast

Seedance 2.0 Fast is the faster, lower-cost variant of Seedance 2.0. It shares the same multimodal capabilities (text-to-video, image-to-video, first-and-last frame, and reference-to-video) but is limited to 480p and 720p output (1080p is not supported).

## Overview Seedance 2.0 Fast is ByteDance’s multimodal video generation model. It generates video and audio together, so sound effects and dialogue timing match the visuals. You describe a scene and the model produces a short, edited-feeling clip with strong motion stability and controllable camera language. This helps teams prototype ads, story beats, and visual concepts without a shoot. ## What you can build - Text-to-video scenes with camera moves like tracking shots, push-ins, and POV switches - Image-to-video animations that start from a provided first frame - First-to-last frame transitions for outfit changes, product reveals, and before/after shots - Reference-driven clips that keep a character or style consistent across a short sequence - Music- or rhythm-guided visuals using reference audio as a timing and mood anchor - Social video loops for Reels, Shorts, and TikTok in vertical or square formats ## Inputs - A scene description written as plain text. Include actions, setting, lighting, and camera direction. - Optional spoken lines written inside double quotes if you want dialogue with lip-synced audio. - Optional first-frame image to anchor the opening shot for image-to-video. Use common web image formats such as JPG, PNG, WebP, GIF, or AVIF. - Optional last-frame image to guide how the clip should end. This works only when you also provide a first frame. - Optional reference image set (up to 9 images) to lock in character appearance, wardrobe, props, or visual style. In your description, call them out as Image 1, Image 2, and so on. - Optional reference audio clips (up to 3 files) in WAV or MP3 to guide mood, pacing, and sound design. Some workflows require at least one visual reference when you include audio. - A target output resolution choice. Higher settings can increase detail, but availability depends on the deployment. - An output aspect ratio choice, or an adaptive option that lets the model pick based on your prompt and references. - A clip duration choice of 5, 10, or 15 seconds. - A toggle to include or omit generated audio. - An optional watermark toggle. - An optional numeric seed to reproduce results. Set it to 0 to randomize. ## Outputs The model returns a rendered MP4 video. If audio is enabled, the MP4 includes a synchronized audio track that can contain ambience, sound effects, music, and dialogue. When you provide reference images, the output typically preserves key identity cues like clothing and overall look within the clip’s length. ## Recommended settings - First tests and prompt iteration: generate 5 seconds at 720p with adaptive aspect ratio. - Dialogue scenes: keep spoken lines short and wrap them in double quotes. - Image-to-video with a specific ending: provide both a first frame and a last frame, then describe the transition. - Character consistency: use 3–9 reference images and explicitly refer to them in the description. - Repeatable results for reviews: reuse the same seed and keep references identical. ## Limitations - Single generations are limited to short clips, with an upper bound of 15 seconds. - Resolution support varies by provider and endpoint. Fast is limited to 480p and 720p (1080p not supported). - Face and likeness controls can be strict. Uploads that include real human faces may be blocked or require authorization. - Requests that reference copyrighted characters, brand IP, or public figures may be filtered or refused. - Reference quality matters. Blurry images, heavy compression, screenshots with UI overlays, or inconsistent reference sets can cause identity drift and visual artifacts. - Audio quality can vary by prompt. Very dense sound instructions can introduce distortion or muddiness. ## Safety & compliance - Only upload images, audio, and prompts you have rights to use. - Don’t generate content that impersonates real people without consent and legal permission. - Avoid prompts that ask for copyrighted characters, trademarked brands, or scenes copied from films. - Use the model for lawful, non-deceptive content. Don’t create misleading deepfakes or fake news footage.