Seedance 2.0 Prompt Writing Guide: Mastering the @ Reference System for AI Video Creation

Seedance 2.0 is ByteDance's multimodal AI video generation model, accepting image, video, audio, and text inputs to produce short videos of 4 to 15 seconds with built-in sound effects and music. Unlike simple text-to-video tools, Seedance 2.0 rewards precise prompts that tell the model exactly what role each uploaded asset plays. Mastering this skill lets creators replicate camera work, reproduce visual effects, extend or edit existing footage, sync visuals to music, and produce commercial, educational, and narrative content efficiently. This guide distills the essential techniques for writing high-quality Seedance 2.0 prompts.

Understanding the input limits is the first step to effective prompting. Seedance 2.0 accepts up to 9 images (jpeg, png, webp, bmp, tiff, gif; each under 30 MB), up to 3 videos (mp4 or mov; each under 50 MB, total duration 2 to 15 seconds), and up to 3 audio files (mp3 or wav; each under 15 MB, total duration up to 15 seconds). The combined file count must not exceed 12. Output duration ranges from 4 to 15 seconds, and resolution spans 480p to 720p. Realistic human faces in images or videos are not supported and will be blocked automatically. Reference videos increase generation cost slightly. Prioritize uploading the assets that most strongly affect the visuals or rhythm, and allocate the file budget thoughtfully.

The @ syntax is the most important element of a Seedance 2.0 prompt. Each uploaded asset is referenced as @图片1, @视频1, or @音频1, and the prompt must explicitly state its purpose. Common roles include first frame or last frame anchoring, character appearance, scene or background, camera movement, motion choreography, special effects and transitions, beat or rhythm, voice timbre, background music, sound effects, clothing, product appearance, and typography. Vague references such as 'refer to @视频1' are ineffective; instead, specify whether you want its camera work, its motion, its effects, or its rhythm. Multiple references can be combined within one prompt, for example assigning a character from @图片1, camera language from @视频1, BGM from @音频1, and location from @图片2 in a single sentence.

A strong prompt follows a layered structure: subject and character setup, then scene and environment, then action and motion, then camera language, then time-segmented description, then transitions or effects, then audio and sound design, and finally style and mood. For videos longer than about 8 seconds, time-segmented prompts give fine control. Divide the timeline into segments such as 0 to 3 seconds for the opening composition and motion, 3 to 6 seconds for development, 6 to 10 seconds for the climax or key action, and 10 to 15 seconds for resolution, freeze frame, or brand text. Each segment should describe what is visible, how the camera moves, and what action occurs, keeping each moment physically plausible and not overloaded with scene changes.

Precise camera vocabulary sharpens results. Basic movements include push-in or slow push (moving toward the subject), pull-back (moving away), pan left or right, tilt up or down, follow shot tracking the subject, orbit or circular shot around the subject, and one-take continuous shots without cuts. Advanced techniques include the Hitchcock dolly zoom for vertigo effects, fisheye ultra-wide distortion, low-angle hero shots, high-angle or bird's-eye views, first-person subjective perspective, rapid whip pans creating motion blur, and mechanical-arm style multi-angle tracking. Shot sizes range from extreme close-up of eyes or lips, to face close-up, medium close-up of head and shoulders, medium shot from waist up, full shot showing the whole figure, and wide or establishing shot revealing the environment. Specifying these terms helps the model interpret intent accurately.

Different creative goals call for different prompt patterns. For character consistency, anchor the subject to a reference image and describe the sequence of actions in detail. For precise camera replication, state which camera effects, expressions, or movements to copy from the reference video and in what order. For creative template or effect replication, specify transitions, the viewpoint switch from third-person to first-person, and the sequence of visual worlds. For video extension, describe the continuation in timed segments and select a generation length matching only the new portion. For video editing, preserve most of the original footage while stating precisely which element changes, such as swapping a character, altering hair color, or adding an object. For beat-synced montages, reference a video for rhythm and list the images to be cut on key frames. For dialogue and voice, provide each speaker's line, emotional tone, and delivery style. For one-take shots, describe a continuous path through multiple locations without cuts. For e-commerce, describe product disassembly, rotation, or 360-degree showcase movements with clean lighting. For science education, use timed segments showing physiological processes with clear visual metaphors. For AI short drama and comic adaptation, reference storyboard panels and specify dialogue, pacing, and sound effects. For video fusion, describe morphing transitions where particles become a subject and that subject dissolves into the next clip.

Appending style modifiers improves output quality. Visual style terms include cinematic texture, film grain, shallow depth of field, 2.35:1 widescreen, 24fps, black-and-white ink painting, anime style, photorealistic rendering, high-saturation neon palettes, cool-warm contrast, 4K medical CGI, and refined CG animation. Mood terms include tense suspense, warm healing, epic grandeur, comedic exaggeration, documentary restraint, dark fantasy, and high-energy xianxia. Audio direction should be explicit: specify background music style, sound effects such as footsteps, crowd noise, or engine sounds, voiceover timbre referenced from a video, and beat-aligned sound design where footsteps, breathing, and fabric rustle sync with the rhythm. Combining visual, mood, and audio instructions produces coherent results.

Several pitfalls reduce output quality. Ambiguous references such as 'follow @视频1' without specifying what to copy create confusion, so always state whether you mean camera work, motion, effects, or rhythm. Conflicting instructions, such as requesting a fixed camera and an orbiting camera in the same segment, cause unstable results. Overloading short durations with too many scenes breaks physical plausibility, so keep 4 to 5 second segments focused. Unassigned assets waste file budget; every uploaded image must be tagged with its purpose. Ignoring audio misses a major quality lever, so always include sound guidance. Mismatched duration occurs when prompt complexity exceeds the selected length. Finally, uploading materials with clearly identifiable real human faces will be blocked, so plan accordingly.

Practical templates accelerate production. A 15-second product advertisement can open with the product spinning in dynamically, move through multi-angle reveals with light sweeps, show real-use application in the middle segment, and end with a held product shot, brand slogan, and music climax, referencing @视频1 for editing style and transitions. A 15-second short drama clip can feature a close-up of a tearful character, a dialogue exchange escalating in emotion, a revelation of evidence with a shift from anger to shock, and sound design of tense piano and crying voices. A dance video can replicate choreography from a reference video onto a character in a reference image over 13 seconds. A scenic beat-synced montage can cut between six landscape images timed to a reference video's rhythm. A xianxia battle sequence can build from a low-angle hero shot, through orbiting cuts and energy blasts, to a final slow-motion leap and landing with dialogue and lingering sound. A 15-second health education short can show a translucent body with a blood vessel that gradually narrows as fat deposits form, with lighting shifting from clean to dim to convey warning.

Building a prompt follows a clear process. First, define the goal: advertisement, short drama, music video, science explainer, or vlog. Second, inventory the assets: which images, videos, and audio files are available. Third, assign each asset a role such as first frame, character reference, or camera reference. Fourth, construct the prompt by setting the subject and scene, using time segments for videos longer than 8 seconds, specifying camera language, adding audio design, and appending style modifiers. Fifth, check constraints: total files at most 12, no realistic faces, and duration within limits. Finally, refine by removing ambiguity and confirming that every @ reference has a clearly stated purpose. Following this workflow consistently yields reliable, high-quality Seedance 2.0 videos.