प्रोडक्शन गाइड

Seedance 2.0 प्रॉम्प्ट: असली ब्रेकडाउन से शॉट और कॉल की योजना

Seedance prompts work best when they describe one generation unit at a time: which shots it covers, how long it runs, and what stays out of the prompt. This guide follows a real rea-video breakdown of a 57-second (0:57) video in which a creator demonstrates making a product ad with AI tools. Its remake plan needs one Seedance 2.0 call, and the sections below show the rules that plan used.

अपडेट

केस साक्ष्य: मूल ब्रेकडाउन रिपोर्ट पढ़ें · केस फ़्रेम और शॉट समय देखें

प्रकाशित टूल रिपोर्ट पर आधारित। निर्माण वर्गीकरण विश्लेषणात्मक निर्णय हैं; सुझाए चरण मूल प्रोजेक्ट या सत्यापित रीमेक नहीं हैं।

Start from shots, not from one prompt

The example video has 28 shots in 57 seconds: 21 live-action, 6 AI-generated and 1 still frame. Compressing the whole video into one prompt would mix production routes that need different assets. The breakdown separates them first, then writes prompts only for the AI shots. Of the 6 AI shots, two (#2 and #3) are an opening preview cut from the same footage as #25 and #26, so only four distinct shots need generating.

One generation unit, one call

The report’s plan reads: 4 AI shots → 1 generation unit → 1 Seedance 2.0 call (its label covers the Mini and Fast variants). Unit G1 covers shots #25–#28, 49.1–57.7 s, about 8.5 seconds, written as one multi-shot prompt with hard cuts between shots; the two preview shots are cut from that result instead of being generated again. The report does not compare engine versions. Per-call limits and multi-shot support change between releases, so check the current limits of the version you use before committing to a call count.

Grouping rules from the real plan

Merge AI shots that sit in the same set and are joined by hard cuts into one generation. Keep each call at 15 seconds or less, and keep a run of hard cuts inside a unit to about 8 seconds, 10 at most. Split at dissolves, fades and set changes; those transitions mark a new generation unit. These are the rules the report states for its plan. Treat the 15-second cap as the report’s planning assumption and re-check it against your engine.

What stays out of the prompt

The example carries 23 subtitle lines and one text label, all laid on in editing, and the report keeps them out of the prompts. Captions, logos and stickers belong in editing, where wording and timing stay under your control. For voice, the plan is to record the spoken part separately and mix it in rather than let the video model generate it, so you can change the script without regenerating picture; the model can still produce action sounds and ambience. For G1 the report suggests feeding a first-frame image (the source frame of shot #25, or a redrawn one), product photos, a character reference and the source clip as a motion reference, and notes that in this mode the first frame has to be passed as an extra reference image. Any current image model can redraw that first frame — MiniMax’s H3 image, for example; what matters is fixing composition before the video call.

What this guide does not claim

The call plan and rules above come from a real published report, but rea-video does not generate videos, and we have not run these prompts through Seedance. Treat the prompts you derive as structured drafts, test one unit first, and adjust duration and motion to what your engine version returns.

पुन: प्रयोग योग्य विश्लेषण टेम्पलेट

Call # / shots included / total duration
Engine version and per-call limit
First-frame reference (if used)
Per shot: subject, framing, action start→end, camera, light
Keep out: captions, logos, stickers, voice-over
Split points: dissolves, fades, set changes

असली रिपोर्ट के साथ अभ्यास

Download the report and open its AI-generation section: one generation unit (G1, shots #25–#28), the grouping rules, the inputs to feed and the full multi-shot prompt in Chinese and English. It is the source for every rule quoted above.

केस खोलें और रिपोर्ट डाउनलोड करें →

अपना वर्कफ़्लो जारी रखें