FLAM

Cast your own model

Drop two to four photographs of one person into the chat and the house develops a four-view digitals sheet from them. It reads the photographs first for free, quotes the sheet before anything moves, marks its own output, and hangs the result on your casting rail.

Drop photographs of a person into the chat and the house casts them as a model you can use in every frame after. It looks at what you dropped before it charges anything, reads the face for one token, quotes the sheet, develops it, and then marks its own output the way it marks yours. The whole thing happens in the conversation — there is no second screen.

Bring your own model

What you get at the end

One image: a digitals sheet, four views of the same person laid out 2×2 — a tight head-and-shoulders front, a three-quarter turn, a full side profile, and the back of the head. That sheet is the model. It hangs on your casting rail and it is what every later develop attaches as its last reference, which is what holds a face across forty frames.

The photographs you brought ride behind it on the same row, in the order you gave them, so you can always see what the sheet was made from.

Where to start it

Two doors, one conversation:

  • Cast your own model on the casting rail — under the board, and leading the board when you have no models yet.
  • Say so in the chat. Drop photographs of a person and the house offers it.

Both arm the same sequence. There is no wizard route to lose your place in.

The photographs to bring

Two to four is the number. One works, eight is the ceiling, and past four the identity tends to drift rather than sharpen — more angles of one face give the house more ways to average it.

What each set wants:

  • One frame where the face fills about half the picture. This is the one that locks the identity. Without it there is nothing to hold to and the sheet composes a face of its own.
  • A full-length frame if you have one. It settles proportion.
  • The same person in all of them, the same haircut, the same era.
  • Even light. Hard side light in one frame and flat light in another gives the sheet two different faces to average.

The house looks first, and it is free

Before a token moves, the house inspects what you dropped through the casting lens — five of its nine checks, the ones that decide whether a face can be cast at all:

What it looks forWhat it says will happenWhat to do
mixed-subjectthe model will not look like one persondrop the odd one out
face-unreadablethe collage will invent a face instead of hersswap in a closer photo where the face is clearly visible
too-smallthe detail will not survive the developswap it for a larger original
part-hiddenthe hidden part will be inventedadd a reference that shows it
light-mismatchher face will change between frameseven the light across all of them

The codes are ours; what you read is the middle column, against thumbnails of your own frames. face-unreadable covers two situations: no photograph is close enough, or every close one has the face covered — sunglasses, a mask, a hand, hair across it.

A finding that blocks holds the whole sequence. The read does not run, the quote does not appear, and nothing is charged while it stands. Replace the frame or drop a different set and the look runs again, still free.

The read — 1 token

Once the set is clear, the house reads the person out of your photographs and fills in a card: gender, age, hair, eyes, skin, and anything else it noticed. That read costs 1 token, and the invitation says so before you drop anything.

The card is yours to correct. What it read is a starting point, not a verdict. Change any field, or type a sentence — "a small scar above the left brow" — and your words go into the sheet verbatim. Gender is always a real question on the card and never something the read answers, because a house that guesses one has already miscast your model.

Whatever the card ends up holding is stored on the model's row, so the next tool that needs it reads it instead of asking you again.

The quote — 9 or 13 tokens

Then a plan card, itemised, and this is the only place money moves:

LineCost
Read their face1 token, already spent
Develop their sheet13 tokens at full size, 9 at the smaller

Full size is the default and it is the right default. A 2×2 sheet at the smaller size gives each panel roughly 512 pixels of face, which is under the floor every model states for a usable person reference. Full size gives each panel about 1024. The smaller one is there, priced honestly, for when you want a look before you commit.

These are the ordinary develop prices — a sheet is one image generation and it costs what one image generation costs. See quality and cost.

The develop

The sheet develops in the conversation like any other frame. You write nothing: the whole instruction is written from your trait card, server-side, and it carries three things it will not bend on — the opening enumerates the features that must not drift between panels, the hair block leads by locking the hair to your reference, and the closing line restates that only the camera angle changes. Each of the three is there because a sheet once drifted without it.

The sheet also picks its own shape. Nothing forces an aspect on it, because a 2×2 squeezed into a portrait phone photo's proportions comes back with panels cropped off the edge.

Then the house marks its own sheet

The judge reads the sheet before you do, wearing a different set of glasses. A digitals sheet has no garment, no logo and no scene, so those questions are withdrawn — asked anyway they come back empty, and empty reads as perfect. What it asks instead:

MarkWhat it means
grid-brokennot laid out as a 2×2
gender-wrongnot the person you asked for
identity-driftthe panels disagree on the face
hair-occludesthe hair covers it
expression-wrongthe expression should be neutral
panel-duplicatethe same view twice

identity-drift is the one mark that belongs to both worlds. On a normal frame it asks whether the face matches your reference; on a sheet it asks whether the four panels agree with each other, which is the only comparison the sheet can actually answer.

grid-broken and gender-wrong weigh the most, at 25 each, because a sheet that is not a 2×2 or not the right person is not usable at all. panel-duplicate weighs 10 — three good views is a thin sheet, not a ruined one.

When the sheet is not them

Say what is off in your own words — "the hair is in front of the ear in the top-right panel" — and it develops again with your sentence carried in, ranked above every other instruction in the sheet. That pass is the ordinary redraw at the ordinary price, quoted on the same card. There is no special path and no free do-over: the house tells you first, you decide whether it is worth another 13 tokens.

Both passes stay. The strip under the sheet reads which one you are looking at.

Hanging it

Approve, name them — the house proposes a name and you keep it or change it — and the sheet walks to your casting rail. It is a model from that moment: castable in a lookbook, in a try-on, anywhere a model goes. Nothing else needs to know it came from your photographs rather than the board.

The rest of the dossier — her range and her build

Her range and her build

The sheet above locks a face and nothing below the shoulders. So once it hangs, the house says so and offers the two sheets that close the gap:

Noa is a face, and nothing below the shoulders. Want her range and her build? Two more sheets — her expressions, and her full length.

Four answers: her range, her build, both, not now. It is optional in the literal sense — the casting above completes exactly as it always did with none of this mentioned, "not now" costs nothing, and you are not asked again in the same session. The offer is also on the model's own call sheet, under Her range and her build, for a model you cast last week or one you said no to at the time. It is there on models you brought and on nothing else.

Each sheet quotes before it moves, at the same price as the first: 13 tokens at full size, 9 at the smaller.

Her range

A second 2×2 of the same person wearing four different expressions — a calm neutral, a closed-mouth smile, an open smile with teeth, and a serious editorial look. It is the answer to a brief that asks for a smile: without it the house has never seen this model smile and has to invent one.

The judge reads it through its own glasses. expression-missing — a panel is not wearing the mood its position was briefed for — exists only on this sheet, and expression-wrong (a panel is smiling on a sheet that asked for neutral) is withdrawn here, because that is the one thing this sheet exists to show. panel-duplicate re-reads as two panels wearing the same expression: every panel is front-facing here by design, so the repeated thing to look for is the mood, never the camera angle. If you send it back, the correction it carries says the same — the four moods again, camera unmoved.

Her build

Two full-length panels side by side, a straight-on front and a full side profile, head to feet. This is the sheet garment drape actually hangs on.

It needs one photograph of the whole person, head to feet, and it checks the ones you already gave before it asks — that look is free, and it happens before any card appears. If none of them qualifies, the house asks for one and nothing has been charged to find out.

Then it reads the body off that photograph for 1 token and hands you the card, filled in, in the words a call sheet carries:

heighthow tall they read in the frame — tall, average height, petite. Never a number: you cannot measure a photograph.
buildshape and proportion — slim with long limbs, athletic, broad shoulders, narrow hips.
presencehow they hold the frame — still and self-contained, open and easy.
posturehow they actually stand — upright, weight even, weight on the left hip.
handsshape of the hands, when they are genuinely visible. Hands are the thing image models get wrong most, and nothing else in FLAM has ever described them.

Every field is yours to correct, in a sentence, before the sheet develops — correct "petite" to whatever you actually call her. A field the read could not see comes back empty and stays empty. An empty field is a correct answer; a plausible guess is a body you have to argue with, and you would not know it was a guess. A head-and-shoulders portrait comes back with all five empty and no figure detected: this read is about proportion, and there is no proportion in a crop.

Where they go

Both sheets join the same row, in order: the collage first — it is still the identity reference every develop locks to, and still the cover — then the dossier sheets, then your original photographs as provenance. The casting words merge onto the model rather than replacing what is there, so a build read can never erase the face read.

Driving it yourself

Through the API the sheet is one submit: generate-character with mode=grid, your traits as one JSON field, and the photographs as ref1, ref2, … in your order. The instruction stays on the server; mode and traits are the whole of what you send. Registering the result is POST /api/toolkit/models/own with the sheet's asset id as collageAssetId and your originals as assetIds.

The dossier is the same submit with mode=expressions or mode=build. The build read is POST /api/toolkit/ai/extract-build — one full-length photograph in, the five casting words out, 1 token, and figureDetected: false with five empty strings when the frame is a crop. Adding a sheet to a model you already have is PATCH /api/toolkit/models/:id with sheetAssetIds and traits; it orders the references for you and merges the traits rather than replacing them. Your own models only.