The Seedance Director's Guide

Field notes from real renders of a real music catalog — a working playbook, not theory.

AI video tools don't reward better wishes; they reward better direction. The difference between generic output and something people rewatch is a scene description that behaves like a shot-listed brief. This guide is the exact system used on a real release catalog.

1. The credit rule that pays for this guide: probe, then strike

Music-video renders are priced per second of audio, and the meter shows the exact total before you commit. So never fire a full song on an untested prompt:

1. Cut a 28–30 second excerpt of your song — a chorus plus one clear vocal line.
2. Render it at 480p with the exact scene description and reference photos you plan to use. That's a ~60-credit rehearsal instead of a four-figure gamble.
3. Judge three things: does the style hold? does the performer stay on-model? does the lip sync land?
4. Fix the description, re-probe if needed, and only then render the full song at high resolution — once.

2. The 5-part scene description template

The scene box is your director's chair (about 2,500 characters). Order matters — front-load what must never be lost:

Part 1 — Style lock (~400 chars)

Declare a visual religion and forbid everything else. Styles age well when they're judged by their own rules — stylization outlasts realism every time.

Flat 2D hand-drawn illustration, [your style: punk zine collage / watercolor
storybook / neon riso print]. Zero 3D, zero CGI, zero photorealism. Hard ink
outlines, limited palette: [3-4 named colors]. Flat lighting, no cast shadows.
Maintain this exact illustrated style in every frame.

Part 2 — World and performer (~500 chars)

Describe your lead so specifically the engine can't drift: silhouette, costume, one signature feature, the world they live in. If you upload reference photos, this text should agree with them.

Part 3 — Arc by song sections (~900 chars)

The engine maps your song's structure — so direct in its language: "in the verses… when the chorus hits… at the bridge… by the final chorus…". Give the video a journey, not a loop: the world should be visibly further along at the end. One central visual metaphor per song beats five ideas fighting.

Part 4 — Camera intent (~300 chars)

Two or three moves, each tied to a musical moment: "slow dolly-in through the verses; hard cuts on every chorus downbeat; one crane-down from the sky as the bridge opens." Every cut should answer why this frame now — aimless coverage is what reads as slop.

Part 5 — Negatives (~200 chars)

No photorealism, no 3D render, no lens flare, no watermark, no on-screen text
or typography, no extra limbs, no close-up realistic hands, no style drift.

3. Camera vocabulary the engine obeys

• dolly in / dolly out — slow pushes sell intimacy or reveal scale
• lateral tracking shot — follows a moving subject, keep "lead room"
• crane down / crane up — heaven-to-stage transitions, great for intros
• rack focus — attention handoff between foreground and background
• held / static shot — used deliberately, a held frame makes the NEXT cut hit harder
• cuts land on the beat — say it explicitly; rhythm-locked editing is the mode's superpower

4. Design around AI's known weaknesses

Strong art direction hides what the model can't do:

• Avoid close-up realistic hands, flowing hair physics, glass and mirror reflections, and long unbroken camera drifts.
• Prefer "limited animation" grammar: held poses, hard cuts, repeating motifs. Frame-to-frame discontinuity becomes a style choice instead of an artifact.
• Bake texture into the style (grain, halftone, paper, print misregistration) — intentional imperfection camouflages temporal noise AND reads as human hands.
• One palette, one lighting logic, the whole video. Lighting drift mid-shot is a top AI tell.

5. Reference photos: your identity anchor

• Up to 3 clear, front-facing portraits keep the same performer in every scene.
• Skip them and the engine casts a performer that matches the voice — sometimes great, always a gamble. Probe both ways once and pick a lane.
• Match your text to your photos: if your reference wears a red jacket, say "red jacket" in Part 2 — agreement between channels is what locks identity.

6. One master, every platform

• Render the master 16:9 at 1080p — it's YouTube-native and the best source for everything else.
• For TikTok/Reels/Shorts, don't crop: design a vertical frame (a styled 9:16 canvas holding your 16:9 film) so nothing is lost and the page itself carries your brand.
• Burn readable lyrics/captions on vertical cuts — caption data consistently shows double-digit watch-time lifts for music content.
• Cut 25–35s vertical teasers that end where the chorus resolves into the intro — loop-friendly clips earn replay retention above 100%.

7. The pre-render checklist

1. Probe excerpt cut (28–30s, the chorus)
2. Scene description complete: style lock → world → arc → camera → negatives
3. Reference photos agree with the text
4. Subtitles OFF (add your own type later) · Lip sync ON for vocal tracks
5. Read the credit meter before pressing Generate
6. Probe at 480p → judge → full render once at 1080p

That's the whole system. It costs about one probe's worth of credits to learn and it compounds on every song you ever render.

© 2026 AI Music Video Guide. Free to share with attribution. Affiliate disclosure: Seedance links in this guide are affiliate links — subscribing through them supports free resources like this at no extra cost to you. Not affiliated with or endorsed by seeddance.video or any AI model vendor.