Ad Lab
Mini product ads as preset recipes: each concept is a structured, deconstructed prompt — aesthetics, beat-by-beat action, text overlay spec, sound design — with the product swappable per SKU. Design the concept once; run any product through it, and edit any part of it when the brief moves.
1Pick a concept
2The recipe — Reverse Rewind
The concept, deconstructed into the four things a video model actually reads. Edit any of it to make the concept yours.
Core aesthetics
- Top-down, overhead flat-lay. Static camera — zero panning or zooming.
- Snappy, high-contrast stop-motion aesthetic.
- Monochromatic ultra-bright single-color background; minimalist environment.
- Reverse-chronological timeline: finished product back to original packaging.
Sound design (rendered by Seedance 2.5 Reference)
- Rewind / reverse-playback whoosh distortion matching the reverse visuals.
- Utensil clinks and soft food movement, played backwards.
- Crisp cash-register ding synced to the price text appearing.
Action sequence
- 1 · The finished product: Fully prepared, plated item centered on the colored background.
- 2 · The removal: A human hand enters frame and pulls the plated item away.
- 3 · Deconstruction — toppings: A utensil scoops the topping off in fluid, reverse-gravity motion.
- 4 · Deconstruction — base: A second utensil lifts the base food upward and out of frame.
- 5 · Raw transformation: Cooked product snaps to raw state, flying backward out of the cooking vessel and sliding into its branded packaging.
- 6 · The product shot: A hand places the raw, packaged product flat against the background.
- 7 · Text overlay: Bold stark text snaps on: brand top-left, tagline middle-left, price bottom-right.
Text overlay
Brand name top-left, tagline middle-left, price bottom-right — bold, stark, snapping on in sync with the final audio cue.
3Your product
The photo fills the fields below (brand read off the pack, a background color reasoned from the packaging) and grounds the video as its first frame.
4Format and cost
720×1280 · 8s · ~$3.70How this is priced — 720×1280 × 8s = ~$3.70
Seedance bills by token, not by second. Tokens ≈ width × height × seconds × 24 ÷ 1024, so pixel area matters as much as length. That is why the shape you pick changes the price as well as the crop: at the same resolution a 21:9 frame has more than twice the pixels of a 1:1 one. 1080p is also billed at a higher rate per token on top of having four times the pixels of 480p. Draft at 480p until the take is right.
5Sound
Native audioSeedance 2.5 Reference: Sound and picture are generated jointly in one latent space, and this endpoint also takes audio IN — a supplied track becomes a timing signal the cuts key off, which is the one thing that makes a beat-driven concept land on purpose instead of by luck.
What layered audio does and doesn't guarantee
- Sound effects sync; music does not. The video model generates SFX from the picture it is making, so hits land on frame. The music model never sees the video — it composes from text alone.
- You get two files, not one. Nothing is muxed: the finished ad exists after you combine the MP4 and the track in an editor.
- The preview is an approximation. Two media elements synced on play/pause/seek — enough to judge the vibe, not sample-accurate.
- The bed is generated 2s longer than the cut, giving you handles to slide a downbeat onto the money moment.
- No mix is applied — no ducking under SFX, no level matching, no mastering.
- Expect variance: 2–4 generations to land one you like.
Spot effects — ElevenLabs Sound Effects v20 of 3 generated · ~$0.006 each
Seedance 2.5 Reference already renders effects from the picture it is making, which is why they land on the right frame — but it approximates a described effect rather than producing it. These are the hero hits: generated exactly as described, delivered as separate files, and placed by you in the edit. One event per generation.
Rewind / reverse-playback whoosh distortion matching the reverse visuals.
Utensil clinks and soft food movement, played backwards.
Crisp cash-register ding synced to the price text appearing.
These lines come from the recipe's sound design — edit them in step 2 to change what gets generated.
How the layers stack, and how to describe an effect
- Native audio is baked into the MP4. It cannot be separated out later — if the model rendered a pour, that pour is in the file. Switch native audio off (Silent) when you want the effects track entirely under your control.
- Spot effects arrive as separate files, one per effect. Nothing is placed for you: you drop each one on its frame in the edit, which is exactly the point — placement is a decision, not a guess.
- Music is a third file, laid under both. On a model that reads audio in, it can also steer the cut — see the timing reference above.
- Layered on top of native audio, a spot effect is a sweetener: it thickens the hit the model already made. Over a silent take it is the whole effects track.
Describing an effect
- Describe the physical event, not the feeling: “the metallic snap of a ring-pull, then carbonation hissing out” beats “refreshing can sound”.
- Name the material — glass, foil, kraft paper, brushed steel. Material is most of what a listener identifies.
- Say how close the mic is. “Close-mic’d, intimate, dry” gives you an ASMR hit; add “in a large room” and you get reverb baked in you cannot remove.
- One event per generation. Two effects in one prompt gives you a muddle of both; generate them separately and place them separately.
6References
0 of 50Reference recipe for Reverse Rewind — what to add, and why
- 1Your product photo, shot straight-on with the label readable.still · Product identityDo this oneBecomes [Image1] and anchors identity. Everything else is judged against it.
- 2A second angle of the same product — three-quarter or side.still · Product identityDo this oneTwo angles give the model geometry to hold onto. This is the single biggest reduction in drift.
- 3A 3-5 second clip whose camera move you want imitated.clip · Motion / cameraCamera language is far easier to show than to describe. Trim tight — the model reads the move, not the content.
- 4A photo of the product's prepared form, if you have one.still · Product identityThis concept ends on the raw pack but opens on the finished dish — showing both halves stops the model inventing the food.
Order matters: references are numbered as you add them, and the prompt addresses them by that number. Add the product first.
Abstract on purpose. The model reads a clip's camera move, cutting rhythm and energy and applies them to your product, so a reference with no subject in it has nothing to leak into the render — only motion.
Orbital drift
Slow orbital motion with soft collisions and long easing. Point a product at this when it should feel weightless and premium — it lengthens every move and removes urgency. The wrong choice for anything cut to a beat.
Pulse grid
Hard rhythmic pulses on a regular interval. This is the one to use when the concept cuts to music — it gives the model an explicit tempo to land actions on, which is the difference between a grid that fills on the beat and one that fills whenever.
Liquid bloom
Fluid expansion outward from centre, dense and organic. Reads as pour, splash and bloom physics. Use it for food and drink where the payoff is something spreading or bursting rather than something moving.
Light sweep
A raking highlight travelling across a dark field. Borrow this for lighting behaviour rather than movement — it teaches the model how a specular hit should travel over a surface, which is most of what makes packaging look expensive.
No starter clips installed. Drop the four files into public/references/ using the names above and they appear here. They are served from this site, so unlike an upload they cost nothing and never expire. Which files exist is read at build time — in dev that is every request, but a production server needs a rebuild to see a newly added clip.
Uploading and pasting a URL are different paths. An upload goes through fal's storage service, which is permissioned separately from generation — so a key that renders video fine can still be refused a file upload. A URL is handed straight to the model, so it needs no upload, no live mode and no storage access. It has to be a direct link to the file (ending in .mp4, .mov, .mp3…) that fal can reach without signing in.
Add more angles of the product to tighten the identity lock, a still to borrow a palette from, a short clip whose camera move and cut rhythm you want imitated, or a track whose beats the action should land on. Each gets a job in the prompt — set it in the dropdown. Trim clips before uploading — these models read the camera move, not the content, so anything past a few seconds costs upload time and buys nothing.
7Compose the prompt
The recipe, your product fields and every reference job are compiled into one prompt, then polished for the target duration. Nothing is generated yet — read it before you spend.
8Generate
Before you spend $3.83
2 high-impact references are missing.
- Your product photo, shot straight-on with the label readable. Without a product reference the model invents the packaging outright.
- A second angle of the same product — three-quarter or side. One angle leaves the model guessing at the sides and back, so the pack warps as the camera moves — the most common reason a take gets paid for twice.
Adding these costs nothing. Generating without them usually costs the price of a second take.
Compose the prompt first.