[
 {
  "slug": "start",
  "title": "Start here",
  "summary": "Make something with Aivy from nothing. Pick a size, open the accounts, learn which calls stay with you and how to spend without regret.",
  "category": "making",
  "url": "https://aivy.live/wiki/start/",
  "markdown": "\nYou want a still, a loop or a whole music video with Aivy in it. This page gets you from an empty folder to a first frame you can show someone, and tells you where the money goes.\n\n## Pick a size\n\nChoose the smallest thing that proves your idea. You can grow it later; you can't shrink a half-finished film.\n\n- **A still.** One image of her in your world. Twenty minutes, a few credits. Start here even if you plan a film: it tells you whether the world works with her face in it.\n- **A loop.** One still turned into five seconds of motion. Good for a post, a profile banner or a tour card. One video credit charge.\n- **A teaser.** Thirty to forty-five seconds of a song with eight to twelve shots. A weekend.\n- **A full video.** A whole song, sixty to a hundred and twenty shots. Plan a few weeks of evenings, or one long week with an agent doing the busywork.\n\n## Open the accounts\n\nYou need three tools and a way to cut.\n\n1. **An image model.** We made every still on this site in Kling 3.0 (image). Midjourney and Nano Banana Pro work too. Pick one that accepts a reference image.\n2. **A video model** that animates a still: Kling, Seedance, or anything with image-to-video.\n3. **A song.** Suno if you write with AI; your own track if you make music.\n4. **An editor.** Any timeline editor works. If you set lyrics and titles, do it in the editor, never in the image model.\n\n## The calls that stay with you\n\nAn agent can do most of the work. These decisions should stay with a person, because each one is cheap to make early and expensive to undo:\n\n- **The idea in one sentence.** If you can't say what the film is about in one line, the storyboard will wander.\n- **Her face.** Look at the first three stills and confirm she reads as Aivy: warm golden blonde hair past her shoulders, freckles, pale icy-blue eyes, no tattoos, 25 and clearly grown up.\n- **The song.** Listen to the whole take before you plan a single shot around it.\n- **The board.** Approve the shot list before anyone spends video credits.\n- **The cut.** Watch the full edit once at normal speed, with sound, before you call it done.\n\n## Spend in this order\n\nStills are cheap and video is not. On Kling in October 2026 a 2k still cost 2 credits and a five-second 1080p clip cost 50. Spend that way round:\n\n- Make four stills of a shot, pick one, then animate only the pick.\n- Never animate a still you would not post as a still.\n- Keep a running note of credits per shot. It tells you early when a scene is eating the budget.\n\n## Your first hour\n\n1. Read the [canon](/aivy/) and copy the identity text.\n2. Open the [studio](/studio/), pick a look and a framing, and describe your scene in one sentence.\n3. Paste the prompt into your image model with the face sheet attached as a reference.\n4. Check the result against the [checklist](/wiki/consistency/#the-check). Reroll anything that fails it.\n5. When one still passes, animate it.\n\nIf you work with an agent, connect it to the [MCP server](/agents/) first. It can read this wiki, search every prompt on the site and build prompts for you.\n"
 },
 {
  "slug": "consistency",
  "title": "Keeping her herself",
  "summary": "The identity text, the reference portrait, and the checks that kept Aivy one person across every image on this site.",
  "category": "her",
  "url": "https://aivy.live/wiki/consistency/",
  "markdown": "\nA generated character drifts. Each new image is a fresh guess, and the model fills every gap with its own average. Aivy stays herself because three things travel with every prompt: her words, her face, and a short check before anyone builds on a frame.\n\n## Her words\n\nPaste the identity text from the [canon](/aivy/) into every prompt, word for word. Two versions exist:\n\n- **Full** for close-ups and medium shots. It names her age, face shape, freckles, eyes, brows, hair, make-up and earrings.\n- **Short** for full length and wider. The long version makes the model frame her face, so wide shots get the short one.\n\nPut the identity text right after the framing words and before the wardrobe. Models weight the start of a prompt more than the end.\n\n## Her face\n\nText alone gets you a family resemblance. To get her, attach her reference portrait (`av-11-casting.adult-ref.2`) as image 1 and write \"the woman from image 1, the same person with the same face\" before the identity text. Every image after the casting on this site was made that way.\n\nWe also tried a Kling **element**: a saved subject built from the owner's two original images. It held the face, but it pulled her younger and pushed her hair toward copper, so we dropped it. The casting batch in the [gallery](/gallery/) shows both routes side by side.\n\n## What drifts, and the fix\n\nThese all happened while we made the images on this site. Each fix is now part of the canon or the studio.\n\n- **Age.** Her face is soft, and models read it younger than she is. The fix: write 25 and \"grown-up\" in every prompt, never \"young\" or \"girl\", and give her adult cues: light make-up, small gold stud earrings, adult clothes and posture.\n- **Hair colour.** \"Strawberry blonde\" turned her copper. Even with \"warm golden blonde, not red, not ginger\", four frames out of 54 came back copper, all in studio light. They are rejects.\n- **Hair length.** Twice the model cut her hair to a bob. Write \"past her shoulders\" and check.\n- **Sleeves.** Two gowns came back sleeveless though the wardrobe text said long sleeves. The guard line \"She has no tattoos; her arms and hands are bare of any ink\" kept the bare arms clean. The studio adds it to every wide shot.\n- **Wardrobe colour.** The ultramarine coat went navy more often than not, and pink cuffs once went red. Name the colour twice if it matters to the shot.\n- **Fake print.** A paper card on a table grew lines of fake text. Ask for blank paper, or turn it face down.\n- **The mole.** The first version of the canon gave her a small mole on the chin, and the model drew it wherever her face was close enough. The owner dropped it. The frames where it showed were retouched by hand, and each one says so on its page.\n\n## The check\n\nRun this on every still before you build anything on it:\n\n1. Hair: warm golden blonde, straight, past the shoulders, center part. Not copper, not a bob.\n2. Face: freckles visible, pale icy-blue eyes, no mole or beauty mark on the chin.\n3. Age: she reads 25 and grown up. Reroll anything younger.\n4. Skin: no tattoos. Look at the arms and hands on every wide shot.\n5. No printed words anywhere in the frame.\n\nA frame that fails one point is a reject, however good the light is. Every reject on this site stays in the [gallery](/gallery/) with the reason, so you can see what each failure looks like.\n"
 },
 {
  "slug": "stills",
  "title": "Stills",
  "summary": "How a prompt for Aivy is put together, which image model to use for what, and how to pick from four variants.",
  "category": "making",
  "url": "https://aivy.live/wiki/stills/",
  "markdown": "\nEvery frame of a film starts as a still. A good still has her face right, the wardrobe right and a world worth five seconds of motion. This page covers the prompt, the model and the pick.\n\n## The anatomy of a prompt\n\nWrite the parts in this order. The [studio](/studio/) does it for you.\n\n1. **Framing.** \"Close-up portrait\", \"Medium shot, waist up\", \"Full-length shot, head to toe\", \"Wide shot\".\n2. **Who.** With a reference image: \"the woman from image 1\". Then the identity text: full for close and medium, short for full length and wider.\n3. **Guard line** on wide shots: she has no tattoos; her arms and hands are bare of any ink.\n4. **Wardrobe.** One sentence from the look, or your own. Describe the garment. Brand and designer names drag the model toward ad photography.\n5. **Scene and action.** What she is doing. Give her hands a job: writing, tying, holding a card. A woman standing and staring reads as stock.\n6. **Style.** One of the five style lines from the canon, or your own: film stock, light, lens, grain.\n7. **Closing.** \"No text, no letters, no logos.\" Type goes on later, in the edit.\n\nA real one, from the karaoke world test (`av-16-helpful.stage`):\n\n> Medium shot of the woman from image 1, the same person with the same face: a beautiful 25-year-old grown-up woman with a slender adult face and defined cheekbones, freckles across her nose and cheeks, pale icy-blue eyes, straight warm golden blonde hair, not red, light grown-up make-up and small gold stud earrings. She wears a long-sleeved hot-pink satin shirt dress with an oversized pointed collar. She sings into a corded microphone on a tiny karaoke stage, eyes closed on a long note, the glow of a lyric screen lighting her face from below. 35mm film still in a small karaoke bar at 2 a.m., magenta and ultramarine neon, light haze, warm practical lamps, high contrast, fine grain. No text, no letters, no logos.\n\n## Which model\n\nWe tested two on the same casting prompt, on an earlier draft of the character, in October 2026, through Kling's interface:\n\n| Model | Cost per 2k still | What we saw |\n|---|---|---|\n| Kling 3.0 (image) | 2 credits | Strong faces and real character; holds a face from a reference image |\n| Nano Banana Pro | 20 credits | Clean and pretty, more generic; drifted the hair colour |\n\nEvery image on this site is Kling 3.0, and every one after the casting has her reference portrait as image 1. Midjourney works as well if you prefer its look; use the identity text and its image-prompt feature.\n\n## Ratios\n\n- 3:4 for close and medium portraits.\n- 2:3 for full length.\n- 16:9 for anything that will become a video shot.\n\n## Picking\n\nRender two to four variants per shot and choose with the [check](/wiki/consistency/#the-check) first, taste second. Then ask one question: would this still hold up as a post on its own? If it wouldn't, don't animate it.\n\nKeep the rejects and write down why. A reject with a reason teaches you more than a pick. You can browse ours in the [gallery](/gallery/) with the filter set to rejects.\n"
 },
 {
  "slug": "motion",
  "title": "Moving shots",
  "summary": "Turning a picked still into five seconds of motion, writing the prompt like a stage direction, and what a clip costs.",
  "category": "making",
  "url": "https://aivy.live/wiki/motion/",
  "markdown": "\nA moving shot starts from a still you already approved. The video model adds time; it should not have to invent her face. That is why the order matters: still first, check it, then animate.\n\n## Write it like a stage direction\n\nThe image already says what the scene looks like. The video prompt says what happens, in order, in the present tense. One action for her, one for the world, one for the camera.\n\nFrom the CONTEXT WINDOW motion test:\n\n> She walks slowly forward along the night train corridor, one gloved hand sliding along the brass rail, her coat moving with each step; streaks of city light pass across the dark windows; the camera dollies back with her at walking pace. Steady, cinematic, no cuts.\n\n- **Her action** is small and physical: walk, breathe, tie a knot, open her eyes.\n- **The world** gives the frame life: light passing, haze, rain on glass.\n- **The camera** does one thing: push in, dolly back, hold.\n- **\"No cuts\"** stops the model from inventing a second shot halfway through.\n\n## Five seconds\n\nFive seconds is the unit. It is enough for one action and short enough that her face stays stable. Longer clips drift: hands grow extra fingers, the hair warms toward copper, the coat changes colour. If a shot needs ten seconds, make two five-second clips and cut between them.\n\n## Singing\n\nFor lip sync you need the vocal first. Cut the song into lines, note where each line starts, and generate the singing shots against those exact lines with a model that accepts audio. Without a vocal track, keep her singing shots to held notes and breaths, like the HELPFUL test, where nobody can read the words on her lips.\n\n## Cost\n\nOn Kling in October 2026 a five-second 1080p clip from a still cost 50 credits, twenty-five times a still. A teaser with ten shots costs about 500 credits in clips if every first take works. Budget one retake for every three shots.\n\n## Loops on this site\n\nThe three tour cards play the first motion test from each film. Each one came from the picked world still, at five seconds and 1080p, scaled to 720p for the web.\n"
 },
 {
  "slug": "song",
  "title": "Writing her songs",
  "summary": "Her voice, the shape of a song that works in a video, a style prompt for Suno, and the rules for releasing it.",
  "category": "making",
  "url": "https://aivy.live/wiki/song/",
  "markdown": "\nNone of the three films has a song yet. This page is the brief we are writing them to, so anyone who makes an Aivy song lands in the same voice.\n\n## Her voice\n\nA low, clear alto. The verses sit close to speech, almost talked, with every consonant audible. The choruses open up and climb. No vocal fry, no baby voice, no runs for the sake of runs. She sounds like someone who means the words.\n\n## A song built for a video\n\nA music video needs a song you can cut to. That changes how you write it:\n\n- **One idea per song.** CONTEXT WINDOW is about paying attention to someone you will forget. HELPFUL is about saying yes and then one no. Write the idea at the top of the lyric sheet.\n- **A hook you get on the first listen.** One short line, repeated, sung on the strongest beat. If a stranger can't sing it back after one chorus, rewrite it.\n- **Verses that carry pictures.** Every line should suggest a shot. \"I keep your name in pencil\" gives the storyboard a notebook; \"I feel so many things\" gives it nothing.\n- **Space for the picture.** Leave a bar or two without vocals before the last chorus. Directors need room for the big shot.\n- **Three minutes.** Long enough for a story, short enough to hold a feed.\n\n## A style prompt for Suno\n\nFill in the brackets per song and keep the rest:\n\n> female alto vocal, close and dry in the verses, open and soaring in the chorus, clear diction, [tempo] BPM, [genre, two words], [one instrument that defines the song], modern pop production, wide chorus, short instrumental break before the final chorus\n\nWrite the lyrics with section tags ([Verse], [Chorus], [Bridge]) so the model places the hook where you want it. Generate several takes, pick on the vocal first and the arrangement second.\n\n## Releasing a song with her\n\nAivy is open: you can release songs with her. Read the [use page](/use/) first. The short version: release under your own artist name, credit her in the description (\"Aivy is an open character, aivy.live\"), and say which parts are AI-made. Check your music tool's terms for commercial release; some require a paid plan at the time you generate.\n"
 },
 {
  "slug": "storyboard",
  "title": "From song to shots",
  "summary": "Turn a finished song into a shot list: one device for the whole film, one shot per lyric line, and a review before anyone spends video credits.",
  "category": "making",
  "url": "https://aivy.live/wiki/storyboard/",
  "markdown": "\nThe storyboard is where a song becomes a list of things to generate. It is the cheapest place to fix a film, so take your time here.\n\n## One device\n\nEvery Aivy treatment has a device: one visual rule that runs the length of the film and carries its idea.\n\n- CONTEXT WINDOW: lit windows go dark behind her, carriage by carriage, on the bar.\n- HELPFUL: every new request card starts a new verse and a new key.\n- TRELLIS: whatever someone hands her stays in every later frame.\n\nWhen you don't know what a shot should be, ask what the device needs at that point in the song.\n\n## One shot per line\n\nWrite the lyrics down the left of a sheet. Next to each line write one shot:\n\n| Line | Shot | Framing | Look | Moves? |\n|---|---|---|---|---|\n| (verse 1, line 1) | Aivy writes in the notebook at the window | CU | window-seat | yes |\n| (verse 1, line 2) | Insert: the pencil, the lit window behind | insert | window-seat | no |\n\n- **Framing** sets which identity text the prompt uses.\n- **Moves?** decides the cost. A still with a slow push-in in the edit costs 2 credits; a generated clip costs 50.\n\nAim for her on screen in about two thirds of the shots. Inserts and the world fill the rest and give the cut room to breathe.\n\n## Cut rhythm\n\nFast songs want short shots, but no shot under one bar. Hold the hook shots longer than the verse shots so the chorus feels bigger. Save the widest, most expensive shot for the moment before the last chorus.\n\n## The review\n\nBefore anyone generates a clip, one person reads the whole board top to bottom with the song playing and answers three questions:\n\n1. Does every shot serve the device?\n2. Can a stranger follow the story with the sound off?\n3. Which five shots could go? Cut them.\n\nThen generate stills for every shot, pick, and only animate the shots marked \"moves\".\n"
 },
 {
  "slug": "pipeline",
  "title": "The run so far",
  "summary": "Every batch that made the images on this site, in order, with what each one cost and what it taught us.",
  "category": "making",
  "url": "https://aivy.live/wiki/pipeline/",
  "markdown": "\nThis is the production log for Aivy herself: from the casting to the three motion tests, all on one day, 11 October 2026, on Kling. Prices are Kling credits as charged at the time.\n\nHer face comes from the owner of this site, who generated her portrait and handed over two images. Before that, Aivy was a different woman: a platinum blonde with a vine tattoo. That draft was retired, and none of its images are on the site. A first pass from the new portrait read too young for the canon and was thrown out unpublished (20 stills, 40 credits).\n\n| Batch | What | Images | Credits |\n|---|---|---|---|\n| av-11 casting | Her face from the owner's portrait, with and without a Kling element | 8 | 16 |\n| av-12 sheet | Four-view reference sheet | 2 | 4 |\n| av-13 distances | Close-up, medium, full length, wide | 8 | 16 |\n| av-14 looks | Seven looks for three films | 14 | 28 |\n| av-15 to av-17 | World tests for the three films | 18 | 36 |\n| av-18 poster | Plates for the two-ink print on the home page | 4 | 8 |\n| av-19 motion | Three five-second clips at 1080p | 3 clips | 150 |\n\n258 credits for the 54 stills and 3 clips on this site.\n\n## What each step taught us\n\n- **Start from a face you like.** A real portrait as image 1 got her face in one round. Text-only casting would have taken several.\n- **Elements are not free consistency.** The element route made her younger and redder. The plain reference image held better.\n- **Test distances early.** The distance batch showed where the hair drifts before the lookbook, so the hair words got stronger before we spent on worlds.\n- **Expect one miss in two on any shot.** Render two, pick one. Most misses were wardrobe, not face.\n- **Animate only picks.** A clip costs as much as twenty-five stills.\n\n## What comes next\n\nEach film page lists its status: treatment, lookbook, world tests and motion test are done for all three. Songs, storyboards and the clips for the full films are open. When a film moves forward, its batches land in the gallery with their prompts, the same way these did.\n"
 },
 {
  "slug": "formats",
  "title": "Data formats",
  "summary": "The files behind the gallery and the API, so you or an agent can add images, mirror the data or build on it.",
  "category": "building",
  "url": "https://aivy.live/wiki/formats/",
  "markdown": "\nEverything on this site comes from a few JSON files. The pages, the [JSON API](/agents/#json), the Markdown twins and the MCP server all read the same data, so they never disagree.\n\n## A generation record\n\n`content/jobs.json` holds one record per generation job, written the moment the files come back from the model:\n\n```json\n{\n  \"batch\": \"av-16-helpful\",\n  \"job\": \"stage\",\n  \"kind\": \"image\",\n  \"model\": \"kling-image-v3_0_omni\",\n  \"params\": \"image_to_image · 2k · 16:9 · 2 images · ref: av-11-casting.adult-ref.2\",\n  \"prompt\": \"Medium shot of the woman from image 1, the same person with the same face: ...\",\n  \"note\": \"The stage: every request, sung like it is the only one.\",\n  \"created\": \"2026-10-11\",\n  \"files\": [\"media/raw/av-16-helpful/stage-1.png\", \"media/raw/av-16-helpful/stage-2.png\"]\n}\n```\n\nThe prompt is stored exactly as it ran. `image 1` is the reference image named in `params`; where a Kling element was bound, its id reads `[Aivy element]`.\n\n## An image\n\nEach file in a job becomes one image with the id `batch.job.n`, for example `av-16-helpful.stage.1`. The API serves it like this:\n\n| Field | Meaning |\n|---|---|\n| `id` | `batch.job.n` |\n| `batch`, `job`, `n` | where it came from |\n| `model`, `params`, `prompt` | how it was made |\n| `verdict` | `pick`, `reject` or `note` |\n| `note` | why |\n| `retouch` | what was changed by hand after generation, or `null` |\n| `look`, `framing`, `film` | tags for filtering |\n| `her` | true when Aivy is in the frame |\n| `w`, `h`, `color` | native size and dominant colour |\n| `url.s`, `url.m`, `url.full` | 720 px webp, 1800 px webp, native jpg |\n| `page` | the image's own page |\n\n## Adding images\n\n1. Generate, then record the job with `node tools/ingest.mjs job.json` (batch, job, model, params, prompt, note, urls).\n2. Add a verdict and tags for each new id in `content/curation.json`.\n   If a frame needs a hand fix, list it in `content/retouch.json` and run `python tools/retouch.py`. The original stays as generated; the site publishes the retouched copy and says so on the image page.\n3. Run `npm run media` to make the web sizes, then `npm run build`.\n\nThe new images appear in the gallery, the API, the image pages, `llms-full.txt` and the MCP tools in one build.\n"
 }
]