Models / Seedance / Start & end frame

Seedance start & end frame prompting guide

How to prompt Seedance in start/end frame mode. Every point links its source; community tips are marked. Your agent gets the same guide through the Atlas MCP (get_guide).

Prompt structure

transition action→camera→audio→constraints

• Formula: precise subject + action details + scene/environment + lighting & colour tone + camera movement + visual style + image quality + constraints. source ↗
• Open with a one-paragraph summary (duration, aspect, one continuous shot, what the first frame looks like and where the last frame lands), then lock subject/world with numbers (8 metres, 85-200mm, golden-hour side sun), then timestamped blocks. community source ↗
• Time-segmented prompts for 10 s+: '0–3s: … 3–6s: … 6–10s: … 10–15s: …'. community source ↗

What the text should describe

• Pass image_url (start) + end_image_url (end); the video transitions from the start image to the ending image. Prompt the motion between them. source ↗
• The ending frame is reached by visible motion (contact, weight transfer, slowing, alignment, hold); budget time for settling and a quiet hold — don't cram a big action into the last half-second or let it snap/dissolve into the end frame. community source ↗
• Resolve contradictions between first and last frame explicitly (door open→closed needs a visible closure; wide→close needs a motivated move or a declared cut). community source ↗

What not to describe

• Don't add reference images/videos/audio in a first+last-frame request (separate scenario on ModelArk). source ↗

Camera vocabulary

• Standard terms work (medium shot, close-up, slow push-in, smooth lateral tracking, fixed shot). Use ONE camera movement per shot; combining push, pull, pan and move raises instability. source ↗

Responds well to

• Name body parts with range, speed and force (slowly raise a hand, quickly turn the head); prefer slow, continuous movement; show emotion through physical detail, not 'very sad'. source ↗

Duration and limits

• fal Seedance 2.0: duration 4–15 s or 'auto'; 480p/720p/1080p/4k; aspect auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16; generate_audio default true. source ↗

Multi-shot

• Complex clips: timeline storyboard 'Shot 1 / Shot 2 / Shot 3', each with who + where + doing what + camera + audio; don't force per-shot durations. source ↗

Avoid and failure modes

• Constraint phrases: 'keep it subtitle-free', 'avoid generating any text or subtitles', 'do not generate a logo', 'do not generate a watermark'. Subtitles cannot be suppressed 100%. source ↗
• Keep dialogue in one language; avoid mixing Chinese and English except proper nouns. source ↗
• Prefer positive mechanisms over long negative lists ('enters through the rear doorway' not 'do not teleport'); keep a few targeted exclusions for observed failures and place each rule next to the event it governs. community source ↗
• Don't ask for 'static camera' and 'orbit' in the same segment; don't overload 4–5 s with many scenes. community source ↗

Negative prompt

Not supported; describe what you want instead. source ↗

Real prompts, credited

An 17-second anime-style short video about a small bookshop on a San Francisco hill, and the afternoon it quietly takes from her. 

How it was made: 6 keyframes → 5 shots, each animated from its start frame (and end frame where needed) in Seedance 2.5 on @higgsfield_ai. 
 
Sample prompt: Hand-painted anime film look, golden-hour light. A young woman in a sage-green knit cardigan stands in a sunlit bookshop aisle reading an open book, then slowly lifts her gaze toward the window with a dreamy smile. Very slow push-in.

Storyboard below ↓

Next

Sources

• docs.byteplus.com/en/docs/modelark/seedance-2-0-prompt-guide
• docs.dev.runwayml.com/api/
• fal.ai/models/bytedance/seedance-2.0/image-to-video/api
• github.com/dexhunter/seedance2-skill/blob/main/SKILL.md
• github.com/mattvideoproductions/MVP-Prompts/blob/main/seedance25_prompting_guide.txt
• seedrouter.ai/docs/seedance-2-0