Seedance 2.0 creators use storyboard grids to plan short-film prompts

• Creators are treating Seedance 2.0 as a shot assembler, with _OAK200's workflow converting a normal storyboard into a grayscale depth map before animation.
• Character consistency is coming from model-sheet packets, with MayorKingAI's griffin short combining GPT Image 2 sheets, a Midjourney environment, Seedance 2.0, Suno, and CapCut.
• Comedy prompts are getting production-board timing, with AllaAisling's Seedance prompt splitting a 15-second fitted-sheet gag into eight 1.9-second beats.
• Depth is becoming a motion-control layer, with CuriousRefuge's test sending Depth Anything output into Seedance 2.0 Omni alongside character and environment refs.

ByteDance's official Seedance 2.0 launch post says the model accepts text, image, audio, and video inputs, including up to 9 images, 3 video clips, and 3 audio clips in one generation. Leonardo's Seedance 2.0 API docs expose an 8-second sample request with motion_has_audio: true, while its Fast docs show start-frame and end-frame guidance. Fal's Depth Anything Video page says it produces temporally consistent per-frame depth maps, including grayscale output and raw .npz export.

_OAK200's revised workflow starts with a normal storyboard, then converts that board into a depth map storyboard before sending it to Seedance 2.0.

The important split is visible in _OAK200's explanation: a regular storyboard carries color, lighting, texture, and style, while the depth version keeps spatial layout, camera position, subject scale, depth relationships, and composition.

The thread's working stack:

1. Generate the normal storyboard first.
2. Convert the board into a grayscale depth map with Nano Banana or GPT Image 2, a route _OAK200 said worked better than asking for a depth-map storyboard directly.
3. Preserve canvas size, aspect ratio, panel layout, divider lines, framing, perspective, composition, and object silhouettes.
4. Use white for the closest surfaces and black for panel dividers, with gray values mapping distance.
5. Send Seedance 2.0 the tone/look reference, the depth map storyboard, and a character sheet if one exists.

The result is less prompt prose about camera and composition, because the board carries that information as geometry.

MayorKingAI's "The Griffin's Flight" thread shows the current model-sheet packet: character sheets first, environment second, video prompt last.

The asset packet had four parts:

• Eryan character sheet, three equal vertical panels: close-up portrait, full-body front, full-body back.
• Griffin character sheet, three equal vertical panels with front, back, wings, talons, beak, feathers, and lion body locked.
• Environment plate, a rocky valley with waterfalls, forest, mist, and paths.
• Seedance 2.0 prompt, with @Image1, @Image2, and @Image3 references for the character, creature, and landscape.

The video prompt also specified natural sound only, ARRI Alexa 65, IMAX-scale anamorphic lenses, golden-hour haze, lateral tracking, slow motion, aerial glide, and exact continuity for faces, wardrobe, feather patterns, proportions, lighting, and landscape.

MayorKingAI's 45-second "KARABO" short used the same packet logic, then turned scene continuity into a reference-management problem.

The first scene used four references: boy, wolf, bear, and snowy forest. The second scene kept those references while staging the bear charge, wolf injury, and spear throw.

For the final scene, MayorKingAI removed the bear reference and added the final frame from scene two as a new reference. That handoff is the useful trick: the next scene inherits the actual ending pose, not just the original character sheets.

The final assembly happened in CapCut, where MayorKingAI added fade-ins, fade-outs, two music tracks, and a GPT Image 2 title card.

AllaAisling's fitted-sheet comedy starts as a fake production board: 15 seconds, 8 shots, 1.9 seconds each, with action labels and director notes.

The Seedance prompt then preserves that timing:

1. 0.0-1.9s, STANDOFF: slow push-in on the man staring down the fitted sheet.
2. 1.9-3.8s, APPROACH: he lifts it like an enemy.
3. 3.8-5.6s, FIRST ATTEMPT: hand close-up on the elastic corners.
4. 5.6-7.5s, BETRAYAL: the sheet springs into his face.
5. 7.5-9.4s, TANGLE: he spins fully wrapped, face poking through.
6. 9.4-11.3s, DEFEAT: he sinks to the floor.
7. 11.3-13.1s, SURRENDER: he crushes the sheet into a ball.
8. 13.1-15.0s, THE STUFF: he shoves it into the closet and holds the twitch.

The gag lands because the storyboard does the pacing before Seedance touches motion.

Magnific ran the branching version of the same idea: ask Claude Fable 5 for four endings, animate each in Seedance 2.0, then let the audience vote.

The shared script rules were compact:

• Include a character sheet, or put the character bible at the top of the prompt.
• Use Seedance negative words. Magnific's examples were "no dialogue," "no lip movement," "wordless," "silent," and "grunts of physical effort only."
• Keep the same 3-act structure: Frame 1 setup from 0-5s, Frame 2 action from 5-10s, Frame 3 ironic reveal from 10-15s.
• End on a freeze frame plus music sting for social video.

That format turns alternate endings into prompt variations, not full rewrites.

CuriousRefuge tested a deeper control workflow for performance transfer: convert reference footage into a depth map, add optional masks or skeletons, then use Seedance 2.0 Omni as the motion renderer.

The five-step version:

1. Start with reference footage, then create character and environment references.
2. Convert the reference footage into a depth map with Depth Anything on Fal.
3. Optionally add colorized character masks with Meta Segment Anything, plus skeleton poses.
4. Feed the depth map into Seedance 2.0 Omni as motion reference with the character and environment refs.
5. Generate the final video.

PurzBeats posted the compressed node-graph version: reference video to depth map, depth map to Seedance 2.0, add reference image, set duration/resolution/aspect, pass audio from the source video, hit run.

Magnific's static-start experiment used three start frames and three shot types: medium wide, medium, and static camera.

The examples were deliberately ordinary:

• Apartment balconies, where a ball slips downward after 10 seconds.
• A snowman in a supermarket, where grabbing a carrot sheds snow onto the produce.
• A balcony watering scene, where a hose kink releases into a burst of water.

The final note from magnific says the videos were generated with Seedance 2.0 and the stills with Seedream 5.0 Pro.

The creator posts scattered Seedance 2.0 across multiple surfaces, not one native app.

• Leonardo: MayorKingAI linked Seedance 2.0 on Leonardo after the griffin workflow, and the KARABO thread linked GPT Image 2 plus Seedance 2.0 on Leonardo.
• InVideo: Anima_Labs used Seedance 2 on InVideo with Agent One for storyboards, shot planning, animation generation, and editing.
• OpenArt Director: AIwithSynthia posted a 2-minute Korean drama short made on OpenArt Director with a Seedance prompt.
• Topview: Artedeingenio used Midjourney and Seedance 2.0 on Topview for horror anime, then linked Topview as the access point.
• Runway: AllaAisling made "Ice Racer" with Seedance 2.0 in Runway.
• Dreamina: pzf_ai made a music video with Seedream 5 Pro for stills and Seedance 2.0 for motion inside Dreamina.
• Firefly Boards to Seedance: icreatelife used Firefly Boards for ideation and visual development, then Seedance 2.0 for the final animation.

The practical pattern is already visible: planning tools make the boards, image models make the references, Seedance animates the shot, and editors stitch the final piece.