"Make a dance video with Seedance" — anyone who's tried knows the pain. The character just stands there, waves their arms awkwardly, or morphs into a completely different person mid-clip.
The fix we found: generate a choreography storyboard with GPT Image first, then feed it into Seedance as a reference. Break a 16-step routine into panels, have GPT draw each pose, and let Seedance animate them in order. The results are surprisingly accurate. Here are 5 prompts that walk through the full workflow.
1. The Choreography Blueprint — 16-Step Dance Sequence
See the full prompt on scenic.sh →
"1. Basic Stance — Stand with your feet shoulder-width apart. Relax your knees. Keep your upper body relaxed. Get ready. 2. Step to the Right — Step to the right. Move your body in the direction of the step. Keep your knees soft... 14. Freeze Pose — Hit a strong pose on the beat. Freeze your body momentarily. Create a powerful shape. 15. Finishing Pose — Finish the dance gracefully... Image 1 as the storyboard sequence reference. Follow the 16 panels in order as a single 15-second animated short"
Why this works: This is the starting point of the entire workflow. Breaking 16 steps into 3-4 lines each gives GPT Image clear instructions for distinct poses per panel. The sequence builds progressively — "Basic Stance → Step Right → Step Left → Body Wave → Hip Sway" — so the difficulty ramps up naturally. Seedance distributes these panels across the timeline to create smooth transitions.
The final line is critical: "Follow the 16 panels in order as a single 15-second animated short" directly links the GPT Image storyboard to Seedance's timeline. Without this instruction, Seedance may only reference the first panel or ignore the sequence entirely.
2. Street Dance Video — Cinematic Full-Body One-Take
See the full prompt on scenic.sh →
"Street Dance Choreography Video (16-Step Sequence) [VISUAL STYLE] Realistic full-color cinematic look. Natural lighting with soft daylight tones, subtle shadows, and a clean, modern street aesthetic. [SETTING] An empty urban street, minimal, clean surroundings with no crowd, no vehicles... [CHARACTER CONSISTENCY] Use image1 as the base identity. The same female dancer appears throughout... [VIDEO STRUCTURE] One continuous shot (no cuts) Full-body framing, camera fixed and centered..."
Why this works: The bracketed section structure — [VISUAL STYLE], [SETTING], [CHARACTER CONSISTENCY], [VIDEO STRUCTURE] — makes Seedance process each instruction independently. The [CHARACTER CONSISTENCY] section with "Use image1 as the base identity" locks the GPT Image character's face and proportions across the entire video.
The combination of "One continuous shot (no cuts)" and "camera fixed and centered" is decisive for dance videos. When the camera moves, Seedance allocates compute to camera motion instead of choreography, reducing movement accuracy. A fixed camera with full-body framing forces Seedance to focus on the dance itself — footwork, weight shifts, and arm paths become significantly more precise.
3. Same Choreography, Different Look — Skirt Physics Simulation
See the full prompt on scenic.sh →
"... [OUTFIT] A stylish fitted top paired with a slightly longer skirt that covers the thighs and extends just above the knees. The skirt should move naturally with the dance, subtle flow during spins and steps. [PHYSICS & REALISM] Cloth physics for skirt movement (gentle sway and rotation response) Natural body mechanics and timing Lighting interacts realistically with the character and ground..."
Why this works: Same 16-step choreography, but with [OUTFIT] and [PHYSICS & REALISM] sections added. Specifying the skirt's length and position — "slightly longer skirt that covers the thighs" — lets Seedance calculate the fabric's weight and inertia. "Cloth physics for skirt movement (gentle sway and rotation response)" determines how far the skirt fans out during spins and how quickly it settles when stopping.
Compare this with prompt #2: same choreography, different outfit and physics settings, completely different mood. The two-step variation technique — change the character's outfit in GPT Image, then add physics properties in Seedance — is how you multiply one choreography into multiple distinct videos.
4. Multimodal Input — Image + Audio Together
See the full prompt on scenic.sh →
"Create img2 that follows the exact sequence and movements from steps 1–16 shown in img1. The music should be aud1. There should be no dialogue, text, or narration."
Why this works: This prompt is two lines long. The brevity isn't the point — the strategy is passing instructions as files instead of text. Load img1 with your GPT Image storyboard, aud1 with a music track, and Seedance references both the visual sequence and the audio beat simultaneously.
"Follows the exact sequence and movements from steps 1–16 shown in img1" replaces a lengthy text choreography description. Longer text prompts mean more information for Seedance to parse, which increases the chance of errors. Showing movements through images eliminates text parsing mistakes and lets Seedance replicate the motions directly. This approach consistently outperforms text-only prompts for complex choreography.
5. Minimal Prompt — Start with One Reference Image
See the full prompt on scenic.sh →
"Create a seamless 15-second dance choreography video based on the uploaded reference image. Use the female dancer from the reference image as the main character"
Why this works: This is your entry point. Generate a single dancer character with GPT Image, upload it, and request "15-second dance choreography" — the simplest possible form. You're letting Seedance handle the choreography details while securing character consistency.
Try this before tackling prompts #1–4. The reference image quality determines about 80% of the result. Your first step is getting a clean output from GPT Image: full body visible, natural pose, clean background. Once this gives satisfying results, add section tags like prompt #2, then design a choreography sequence like prompt #1 to level up the detail.
The GPT Image + Seedance Workflow
- Generate your character with GPT Image — full body, clear face, clean background
- Create a choreography storyboard — break 16 steps into panels using GPT Image
- Feed references into Seedance — combine images with section-tagged prompts
- Vary outfit and physics — same choreography, different moods
- Add audio — sync to music with file-based input
GPT Image handles "what it looks like." Seedance handles "how it moves." Separating these roles lets each tool do what it does best, and the results are dramatically better than trying to do everything in one step. Copy these prompts on scenic.sh and start experimenting.
