A short film, designed on paper before a single shot moved
An unfinished noir comedy about a burglar who picks the wrong house, because the smart speaker on the side table has opinions. Fourteen shots exist as approved stills, five of them animate, and there is no cut yet. It is on this page precisely because it is unfinished: it shows the order the work has to happen in, and it shows what the assembly still costs after every shot looks right on its own.
Stage 1 - design the cast before designing any shot
Nothing was generated shot-first. The cast was designed first as reference sheets: the burglar from four angles, the speaker as a product turnaround with its four LED states, and the two props he steals. Those sheets are the only thing standing between a fourteen-shot film and fourteen slightly different burglars.
Stage 2 - approve a still, then and only then animate it
Every shot is then generated as a still, with the sheets supplied as references and one floorplan holding the geography of the house. A still is cheap to judge and cheap to redo. Only once a still is approved does it get animated, so no motion is ever paid for on a frame nobody signed off.
Step by step
- 01
Draw the burglar before you need him
Marcus is generated once, as a reference sheet rather than as a picture: front, three-quarter and profile, same build, same mask, same jeans, same worn trainers. Nothing in this sheet is a shot from the film. It exists so that every later shot has one thing to point at when it asks what he looks like, which is the only reason he stays the same person across fourteen of them.
The prompt for this stepShow promptHide
{ "task": "Generate a clean design reference sheet for an animated short film (multiple angles or states, side by side).", "style": "Stylized noir GRAPHIC NOVEL illustration: flat inked cel-shading, bold black ink outlines, cross-hatched shadows, high-contrast ink wash, hand-drawn 2D comic-book art in the vein of Sin City and Into the Spider-Verse; limited cold blue-grey palette. The only saturated color in any frame is a single glowing accent (the smart speaker's LED ring, or a device's RGB strip); render no brand names or logo text on any object unless the prompt explicitly asks for it.", "subject": "Character design reference sheet. Marcus, a 34-year-old male burglar of East Asian descent. Wearing a matte black ski mask with frayed edges around eye and mouth holes, dark charcoal gray hoodie, slim-fit dark navy jeans, black tactical gloves, scuffed black canvas sneakers. Multiple angles: front view, 3/4 view, side profile. Athletic lean build, 5'10\". Eyes show wariness and dry intelligence through the mask.", "background": "Presented on a completely flat, solid, uniform studio background of EXACTLY hex #B3B3B3 (RGB 179, 179, 179) - a light-mid neutral gray, never dark, never charcoal, never black, never gradient; no environment, no scenery, no props, no clutter, no shadows cast on the backdrop.", "art_style_reference": "One attached image is an ART-STYLE reference only: match its inked graphic-novel rendering exactly; ignore its background and subject content.", "constraints": [ "solid flat studio background of EXACTLY #B3B3B3 (RGB 179, 179, 179), never dark or charcoal", "no scenery or extra props", "the subject is identical across every angle/state" ] }Three views, one character. Generated in a single pass. - 02
Design the co-star as a product, not as a character
The speaker is the other lead and it has no face, so it was designed the way a real product is designed: one turnaround, four states, each state labelled with the emotion it plays. Blue is neutral, red is angry, green is scheming, amber is smug. Doing it this way means a later shot can ask for the red state by name rather than describing anger to a model and hoping.
The prompt for this stepShow promptHide
{ "task": "Generate a clean design reference sheet for an animated short film (multiple angles or states, side by side).", "style": "Stylized noir GRAPHIC NOVEL illustration: flat inked cel-shading, bold black ink outlines, cross-hatched shadows, high-contrast ink wash, hand-drawn 2D comic-book art in the vein of Sin City and Into the Spider-Verse; limited cold blue-grey palette. The only saturated color in any frame is a single glowing accent (the smart speaker's LED ring, or a device's RGB strip); render no brand names or logo text on any object unless the prompt explicitly asks for it.", "subject": "NORA, a cute, friendly smart-speaker character: a stout, softly rounded matte-black cylinder with a mesh speaker grille on the lower half, a recessed LED light ring near the top rim, and a small round OLED screen on the front showing big, expressive, adorable pixelated 8-bit eyes (eyes only, never a mouth). Four LED + eye states shown side by side: (1) blue ring, calm and content; (2) red ring, grumpy and annoyed but still cute; (3) green ring, sly and mischievous; (4) gold ring, happy and triumphant. Include clear dimension callouts on the sheet for BOTH height and width.", "background": "Presented on a completely flat, solid, uniform studio background of EXACTLY hex #B3B3B3 (RGB 179, 179, 179) - a light-mid neutral gray, never dark, never charcoal, never black, never gradient; no environment, no scenery, no props, no clutter, no shadows cast on the backdrop.", "dimensions": "NORA is a cute, stout, matte-black cylinder, EXACTLY 6.5 inches (16.51 cm) tall and 3.5 inches (8.89 cm) wide - a thick, softly rounded, endearing speaker, never a slender bottle or thermos and never squat; these exact dimensions are identical in every shot.", "art_style_reference": "One attached image is an ART-STYLE reference only: match its inked graphic-novel rendering exactly; ignore its background and subject content.", "constraints": [ "solid flat studio background of EXACTLY #B3B3B3 (RGB 179, 179, 179), never dark or charcoal", "no scenery or extra props", "the subject is identical across every angle/state" ] }One device, four states, and every label on the sheet spelled correctly. - 03
Do the same for anything the plot touches
The laptop and the jewellery bag get their own references for the same reason the cast does: they appear in more than one shot, and an object that changes between two shots reads as a continuity error rather than as a generated image. This is unglamorous and it is most of the work.
A prop reference. Cheap to make, and it stops an argument later. - 04
Generate each shot as a still, with the sheets attached
Each shot is written out in plain language and generated as a single frame, with the sheets supplied as references and a floorplan of the house supplied as another so that doors and rooms stay where they were. What comes back is judged like a storyboard panel: does it match the sheets, does it sit in the right room, is the light continuous with the shot before it.
The prompt for this stepShow promptHide
{ "task": "Generate ONE cinematic still frame from a short film. Every field below is a hard specification; the attached reference images are ground truth for identity, design, and layout.", "style": { "look": "Stylized noir GRAPHIC NOVEL illustration: flat inked cel-shading, bold black ink outlines, cross-hatched shadows, high-contrast ink wash, hand-drawn 2D comic-book art in the vein of Sin City and Into the Spider-Verse; limited cold blue-grey palette. The only saturated color in any frame is a single glowing accent (the smart speaker's LED ring, or a device's RGB strip); render no brand names or logo text on any object unless the prompt explicitly asks for it.", "palette": "Cold blue-tinted lighting, grey-blue world. The only saturated color in any frame is the LED ring of the smart speaker.", "format": "Cinematic 16:9 film still." }, "shot": { "id": "S12", "action": "Wide tracking shot FACING Marcus from the front as he walks TOWARD camera down the snowy street, leaving the crime scene. He wears his black ski mask, hood up. The robbed two-story house (matching the attached exterior master, every window dark and unlit) sits small and receding in the BACKGROUND BEHIND him, getting further away. Tucked under his left arm is NORA, the cute stout speaker (6.5 inches tall, 3.5 inches wide, softly rounded), her gold ring glowing; under his right arm, a thin laptop. Snow falls gently. His breath fogs in the cold." }, "characters": { "marcus": { "who": "Marcus, 34, male burglar of East Asian descent, athletic lean build", "wardrobe": "matte black ski mask with frayed edges around the eye and mouth holes, dark charcoal gray hoodie, slim-fit dark navy jeans, black tactical gloves, scuffed black canvas sneakers", "rule": "the ski mask stays on in every shot, covering the whole head and neck; only his eyes and mouth are visible through its holes", "color_rule": "his clothing stays DARK in every lighting condition: charcoal hoodie and dark navy jeans even under street lamps; never light grey, never washed-out blue" }, "nora": { "what": "NORA, a CUTE, friendly smart-speaker character: an UPRIGHT rounded matte-black cylinder that is clearly TALLER than it is wide (height about 1.85x the width), with a soft rounded top and gently rounded edges, endearing but never squat; a mesh speaker grille on the lower half, a recessed LED light ring near the top rim, and a small round OLED screen on the front face showing big, expressive, adorable pixelated eyes", "size": "an upright rounded cylinder EXACTLY 6.5 inches (16.51 cm) tall and 3.5 inches (8.89 cm) wide - a height-to-width ratio of about 1.85 to 1, her height nearly DOUBLE her width; a cute, softly rounded, clearly TALL speaker, never squat, short, wide, or can-shaped, and never a thin bottle or thermos; exactly matching the attached reference sheet, identical in every shot", "finish": "matte black, never glossy or metallic", "charm": "reads as cute and likeable in every state - even the grumpy red state is endearing, never menacing or industrial", "led_ring": "intensely saturated emissive gold; the single strongest color in the frame, casting a visible gold glow on nearby surfaces", "oled_screen": "pixelated 8-bit EYES only, content; the screen never shows a mouth, only two eyes" } }, "set": { "where": "the established quiet suburban street at 2AM, December, matching the attached reference", "layout": "modest two-story wooden house with a porch, EVERY window dark and unlit (the house is empty); fresh snow on the ground and roofs; bare trees; one street lamp; neighboring homes with faint holiday lights far away", "dressing": { "rule": "empty street, no people, no cars, no city skyline" } }, "constraints": [ "this is one continuous film; every detail must match the attached reference images", "if a top-down FLOOR PLAN image is among the references, use it to place architectural elements (closet, sofa, archway, fridge, windows) in the correct positions and perspective", "no text, captions, or watermarks in the frame (title card excepted); no legible digits or clock faces anywhere; any clock or display in the background stays a soft unreadable glow", "no extra people anywhere", "CAMERA IS IN FRONT of him: he walks toward the lens, the robbed house shrinking behind him. He is unmistakably LEAVING, never approaching a house. No lit house ahead of him / between him and camera.", "the object under his left arm is clearly the smart speaker from the reference, not a lantern or lamp", "the house in the background matches the exterior master exactly", "NORA PROPORTIONS ARE CRITICAL: she is a TALL rounded cylinder, clearly TALLER than she is wide - her height is nearly DOUBLE her width (about 1.85 to 1, i.e. 6.5 inches tall by 3.5 inches wide), matching the NORA reference sheet exactly; never a squat, short, wide, or can-shaped speaker", "EVERY window of EVERY house in the background is DARK and unlit - no warm or glowing window anywhere; only faint far-off holiday lights on distant neighboring homes are allowed" ] }One shot as a still, built from the sheets rather than from a description. - 05
Only animate what has already been approved
The approved still becomes the first frame and the shot is animated from there. Doing it in this order means a shot that is wrong gets caught while it is still one cheap picture, and the motion is only ever paid for on a frame that has already been signed off. Five of the fourteen shots have reached this point.
The close-up, moving. The still it started from was approved first. - 06
What is honestly still missing
Nine shots have a still and no motion. There is no cut, no sound, no grade, and no title sequence. Everything on this page is the part these tools are good at, and the part that is left is the part they do not do for you: deciding the order, deciding the length of each shot, and deciding whether the whole thing is actually funny. That work has not been done, which is why this is on the page as a breakdown and not as a film.
The walk-off, animated but not approved. Watch what the bag does.
The exact prompts
Seedance 2.0 - the videoShow promptHide
SUBJECT: the masked burglar from the input frame - matte-black ski mask, charcoal hoodie with the hood up, dark navy jeans, black tactical gloves - walks slowly toward the camera down the snowy 2AM street, leaving the crime scene. Under his LEFT arm he cradles NORA, the cute UPRIGHT smart speaker (clearly taller than wide, ~1.85:1), her LED ring glowing warm GOLD with two adorable pixelated 8-bit eyes; under his RIGHT arm he carries a slim dark laptop flat against his side. MOTION: one deliberate, unhurried stride after another, his weight rocking gently side to side; he never speeds up or slows, calm and in control. CAMERA: a single slow, smooth retreating dolly that faces him head-on and backs away just ahead of him, keeping him centered and roughly life-size the whole time; over the eight seconds the two-story wooden house on the right visibly RECEDES and shrinks into the background behind him, and the move eases to a gentle settle at the end. One locked continuous take - NO cuts, NO jump to another angle, NO extra camera moves. ATMOSPHERE: fine snow drifts down continuously through the frame; his breath fogs in the cold with each exhale; the gold glow from the speaker under his arm sways with every step and spills warm light onto his glove and the underside of his masked chin; the lone street lamp pools cold blue light; bare snow-laden trees and dark houses line the street. PRESERVE EXACTLY as in the input frame: the character and mask, the wardrobe, NORA (upright, taller-than-wide, gold ring, pixel eyes), the laptop, the house design, the framing, and the inked graphic-novel noir look; cold blue-grey palette with the gold glow the ONLY warm color; every house window stays dark and unlit. AUDIO (ambient/diegetic only, NO speech): the rhythmic crunch of boots compressing fresh snow in time with his steps; a soft, steady cold winter wind; his quiet breathing in the cold; a faint low electronic hum from the speaker under his arm; distant hushed winter-night stillness. NO music, NO dialogue, NO voices, NO singing. No subtitles, no captions, no on-screen text of any kind.
How it was wired
Built as a stage-per-shot Studio graph: the character and device sheets into the reference image inputs, the floorplan into another, the written shot description into the prompt. Approved stills then go to an image-to-video node as the first frame. Five shots have moved. Nine have not.