Seedance 2.0 from ByteDance is more than a way to turn a description into a single clip. Think of it as a directing tool: one generation can contain a complete scene, with different camera angles, changes in shot size, and a short edited sequence. The key is to make the shot boundaries clear.
Camera instructions are not decorative keywords. They are a separate layer of direction, and they need a purpose.
Start with a clear structure: subject → action → environment → camera → lighting → style. Use ordinary filmmaking language, such as dolly-in, orbit, whip pan, and rack focus. If you want several shots, explicitly describe where one ends and the next begins. Bracketed commands, supposedly magic durations, and rigid lists of camera keywords are not a substitute for that structure.
If you remember only three things, make them these:
The following templates and examples show how to turn those principles into a deliberately staged sequence rather than a loosely assembled AI clip.
Sign in to leave a comment.
A practical guide for AI filmmakers: structure cinematic prompts, plan multiple shots in a single generation, keep characters consistent, and separate useful techniques from prompting myths.
Video adds a second layer to an image prompt: movement, both of the subject and the camera. A multi-shot sequence adds a third: editing, or how the view changes over time. That makes the order and clarity of your instructions especially important.
A useful single-shot template:
[Subject + action + scene]. [Camera movement]. [Lighting]. [Style]. Avoid: [...].| Element | What to include |
|---|---|
| Subject | Appearance, wardrobe, and distinctive visual details. Specific descriptions help keep the character recognizable. |
| Action | Clear verbs and intensity: slowly, sharply, suddenly. Describe what gives the shot its movement. |
| Environment | Location, atmosphere, weather, and time of day. |
| Camera | One main instruction per shot: direction, speed, and starting and ending position. Put it after the subject, action, and scene. |
| Lighting | Source, direction, and mood. |
| Style | Cinematic references, color grade, and visual character, in a shared line at the end. |
| Constraints | For example: Avoid: text, watermark, distorted hands, extra fingers. |
Be cautious with rules such as “subject and motion are all you need” or “the first 20–30 words carry the most weight.” The research behind the original guide did not substantiate these claims. Describe the important layers of the shot instead of relying on a word-count trick.
One reason to learn Seedance's prompting approach is the ability to request a sequence rather than a single view: wide shot → medium shot → close-up → pull-back. The approaches below suit different kinds of transitions and can be combined.
Divide the clip into segments and give each segment its own shot size and one main camera movement:
0-3s — Wide establishing shot, slow dolly-in
3-6s — Medium tracking shot from the side
6-10s — Close-up, slow push-in
10-15s — Over-the-shoulder, staticExplicit ranges such as 0–3s make the intended timing easier to read. Treat them as direction for the sequence, not a promise of frame-accurate editing. Keep each segment focused rather than packing two competing camera moves into it.
For a punchier sequence with clear cuts, number the shots and name the transition:
Shot 1 — Wide establishing: the car on the road. Camera: crane up, slow reveal.
Hard cut to Shot 2 — Low-angle tracking: wheels and asphalt. Camera: tracking from the side.
Hard cut to Shot 3 — Close-up: driver's hand on the gear stick. Camera: slow push-in.A common multi-shot mistake is leaving the cuts implicit. Separate paragraphs alone can be interpreted as one continuous take. If you need distinct shots, use “Shot 1 / Shot 2 / Shot 3” and explicitly ask for “Hard cut to.”
For a more fluid progression within the scene:
Tracking shot from the side, then whip pan to a wide establishing shot,
then an orbit shot around the subject.This describes a softer progression than a series of hard cuts. For clearer organization, combine “then” with numbered shots.
Keep the sequence manageable: the original guide recommends roughly 4–5 angles within a clip of up to about 15 seconds as a practical planning guideline. More shots can be skipped, compressed, or become less consistent. A few well-executed views are more useful than seven rushed ones.
You do not need special symbols to describe filmmaking choices. Use natural English and recognizable cinematic terms, then explain what the movement should do in the scene.
Camera movement and focus:
Viewpoints and angles:
First-person POV; bird's-eye or overhead; low angle for a sense of power; high angle for vulnerability; Dutch angle for tension; over-the-shoulder (OTS); macro shot.
Framing and editing:
Frame-within-frame; ultra-wide angle; match cut based on a shared shape or movement.
Specific movement is more useful than technical numbers. Describe direction, speed, and the starting and ending view: not just “dolly-in,” but “slow dolly-in from a low angle, ending on his face.” For the look of the image, describe results such as “shallow depth of field” and “soft bokeh” rather than relying on f/2.8, ISO 800, 24fps, or 85mm to produce an exact result.
A recurring difficulty in multi-shot generation is character drift: the person, clothing, or setting changes from one view to the next. Start with a direct consistency instruction, even when you have no reference files:
Maintain the same character, wardrobe and location across all shots.When using reference files, name them in the prompt and assign each one a role:
Named references and explicit roles give you a concrete way to communicate which details should carry across shots.
Do not build a workflow around unsupported figures such as “reference strength 70–80%,” “a maximum of two references,” or “reducing six images to two cuts drift by 60%.” Those claims were not substantiated in the research behind the original guide. Focus on reference roles and clear visual descriptions instead.
An attached image can be a strong visual anchor, but a reference image and an opening frame are not the same instruction.
Use a reference to define the character's identity, such as face, hairstyle, and clothing, or to establish the style and location you want to retain.
Reference @Image1 = the character's face and outfit.
A medium shot of them walking into a marble lobby, keeping the face and outfit from @Image1.
Hard cut to a close-up (@Image1 for likeness). ...
Maintain identical facial features and wardrobe across all shots.The distinction is important: the reference anchors identity and style, not necessarily the composition of every shot. You can request new views around those shared details.
In image-to-video mode, the image is the actual opening frame. Your prompt describes how the scene begins to move from that starting point.
Starting from the given frame: [subject, scene, lighting and mood visible in the image].
[Subject] begins to [action]. Camera: [move, direction, speed].
Keep the same lighting, wardrobe and composition as the opening frame.Do not contradict the opening image. Keep the initial time of day, wardrobe, and composition consistent with what is already there. Concentrate on what starts moving and how the camera responds, rather than describing a different starting shot.
For either approach, read the image carefully:
Use visible details rather than broad adjectives. Do not add information that is not present in the image.
One useful workflow is to build the opening image in Nano Banana, getting the face, lighting, and composition where you want them, then use that image as the first frame in Seedance. This lets you plan appearance before introducing movement.
The original guide's source review challenged several frequently repeated rules. Treat the following as unsubstantiated shortcuts, not a reliable command language.
| Claim | A more useful approach |
|---|---|
| Bracketed transitions such as [Shot 1: Wide], [Cut to: ECU], [Dissolve to], or [Time-lapse transition] are special commands. | Do not rely on the brackets. Use numbered shots and an explicit “Hard cut to” when you want a cut. |
| There is a fixed set of eight or nine camera movements with mandatory keywords. | Use ordinary filmmaking vocabulary and explain the intended movement. |
| 4 / 5 / 6 / 8 / 10 / 12 / 15 seconds is a universally supported set of durations. | Choose from the current interface. Timing written in a prompt is not proof that a duration is selectable. |
| Clips under eight seconds cause a 500 error, or multi-shot generation only works from 12 seconds onward. | Do not treat these thresholds as universal rules; they were not supported by the original review. |
| Reference strength must be 70–80%, only two references are allowed, or a specific reference reduction guarantees 60% less drift. | Avoid unsupported percentages and fixed recipes. Name references and give them clear roles. |
| Exposure and lens settings alone give precise visual control. | Describe the visible effect, such as shallow depth of field. |
| Separate paragraphs automatically produce separate shots. | Mark the shots and transitions explicitly. |
| A longer prompt is always better. | Prioritize structure and useful detail over length. |
If a client brief or tutorial asks for one of these shortcuts, explain the uncertainty and suggest a clearer, shot-based alternative.
A. Brief to prompt. Identify the subject, action, location, mood, format—narrative, advertising, music video, or art—and whether you need multiple angles. Build the prompt using the scene structure above, then add time ranges or numbered shots if the view needs to change.
B. Reference image to prompt. Read the appearance, wardrobe, location, and palette from the image. Refer to @Image1, connect the relevant details in words, and design new views around them without forcing every shot to copy the reference composition.
C. First frame to prompt. Describe what is actually in the frame, then focus on the movement and camera behavior that develop from it.
D. Product scene. A useful four-shot progression is ultra-wide establishing shot → macro texture detail → close-up of a detail or logo → pull-back reveal. Match the grade and pacing to the brand.
For a story longer than about 15 seconds or more than five shots, do not try to squeeze everything into one generation. Split it into scenes, generate them separately, and keep a shared character reference for a more controlled result.
Neylora's AI prompting assistant follows these principles to help you build a structured prompt. Try the Seedance generator.
Seedance 2.0 rewards the same habits as deliberate filmmaking: clear structure and purposeful camera direction. Build the scene first, make shot boundaries explicit, and keep the sequence realistic in scope. The aim is not to pile on “cinematic” keywords, but to give every shot a clear role in the scene.
You do not need a magic syntax. You need a readable plan for the subject, the camera, and the cut.
Adapted for Neylora Academy from the original Seedance 2.0 prompting guide. The original article describes a review of 17 sources, with 10 claims supported and 15 prompting myths challenged. This English edition is a translation and editorial adaptation, not a new verification of those results. Models and interfaces change quickly; check current controls and test the techniques on recent generations.