+ Loomwright
Independent. A style is built above every book and cares nothing about how the book was written.
A finished bookLookBook
Contact for pricingProduces
A published casting style
LookBook turns reference material into a casting style: a trained LoRA and the four prompt blocks that wrap a subject in it. A style is built once, above every book, and every book in your organisation can be cast in it afterwards.
Requires Aardvark Studio Core (free) · Release 1.0.0
- Who it is for
- A creator or small studio that wants every title to look deliberate, and to be able to say what made the look.
- What it takes from the Core
- Nothing about a book. A style needs source material, not a manuscript — which is why the library lives above the project tree and can be built with no book open at all.
- What you have to run
- The Aardvark Studio server · A GPU render host · A LoRA training host
The change
Before, and after
Today, without it
- A look you can describe but not reproduce.
- Prompt fragments pasted between projects, drifting a little each time.
- A LoRA file somewhere on a drive, and no record of what it was trained on.
- A catalog entry naming a checkpoint that is not actually on the render host.
With LookBook
- A style with its whole provenance on one trail — every source, every frame, every decision.
- The same look available to every book in the organisation, and to none of them by accident.
- One book-scoped decision — which look this book is cast in — and nothing else to remember.
- A published style that could not have published if its checkpoint were missing.
How it goes
A style starts with material and ends with a catalog entry that has been checked. In between, eleven steps: five preview frames so the picker is a choice made by eye, then source, extract, select, review — the gate — build, caption, package, train, define and publish.
Two of those steps exist because a trained LoRA on its own is not a style. Define authors the four prompt blocks that wrap a subject in the look, per family, against live preview renders. Publish confirms the checkpoint is present on the render host by listing the host’s own model directory, and refuses the style if it is not. That refusal is the feature: a catalog that can only name files that exist is worth more than one that names everything you trained.
Where it stops for you
3 gates, and who holds each
- 1
Review
Cinematographer
The gate. Eyes on the frame pool. No heuristic replaces this step.
- 2
Define
Cinematographer
The four prompt blocks are authored per family, against live preview renders.
- 3
Publish
Cinematographer
The checkpoint is confirmed present by listing the render host's own model directory. A style that trained but never reached the host is refused.
The whole pipeline on one trail
Source, extract, select, review, build, caption, package, train, define, publish — eleven rungs with per-step provenance you can hover rather than go and look for. Everything is safe to re-run, and each step's own help says what a second run does with what the first left behind.
- Video clusters, uploaded image sets, or an archive query as the source
- Decode, crop, score and stratify into a balanced frame pool
- Farthest-point sampling thins the pool to a visually varied working set
Screenshot slot
The Trail — a style from source to published, with per-step provenance
A review gate that cannot be automated away
One step in the middle is a person looking at frames. It is called the gate in the product's own documentation for the same reason it is here — no scoring heuristic decides what a style is made of.
- Keeps and rejects recorded per frame
- The dataset is built from what survived, not from what scored well
Screenshot slot
Corpus review — the gate, with keeps and rejects
A style is more than a checkpoint
A trained LoRA is a file. A casting style is a file plus the four prompt blocks that wrap a subject in it, authored per family against live preview renders — which is what makes it something VisualBook can actually render with.
- Four prompt blocks, authored per family
- Live preview renders while you write them
- Style bench and look tests before anything publishes
Screenshot slot
Style definition — the four prompt blocks against a live preview
Publishing verifies rather than asserts
A style becomes selectable only after the trained checkpoint is confirmed present by listing the render host's own model directory. A checkpoint that trained but never reached the host is refused with a sentence saying so, instead of a catalog entry pointing at a file that is not there.
- The check reads the render host, not a database column
- The refusal names what is missing
Org-scoped, with one book-scoped decision
A published style is the same look in every project and with no project open at all, so the library lives in a rail container rather than in a book's tree. The single book-scoped surface carries one question — which look is this book cast in.
- The library is above every book
- Pre-Production ▸ Cinematography holds the per-book choice
Sample output
What comes out of it
These are slots. No sample from a real title has been cleared for publication yet, and a mock-up of one would be the exact thing this page is trying not to do.
Screenshot slot
The Trail — a style from source to published, with per-step provenance
Screenshot slot
Corpus review — the gate, with keeps and rejects
Screenshot slot
Style definition — the four prompt blocks against a live preview
What it needs
Be clear about the hardware
- required
The Aardvark Studio server
One container, deployed once. Every product needs it.
- required
A GPU render host
Preview renders during definition, and the host whose model directory publish reads to confirm the checkpoint actually arrived.
- required
A LoRA training host
Training is the step that turns a dataset into a style. It runs on a GPU you provide.
- optional
A captioning service
Replaces caption stubs with what each image actually shows. Without it, captioning is manual.
Roles it adds
Who can do what
Cinematographer
Builds styles. Deliberately not the role that casts a book in one — see the product's own note on why those are two jobs.
Worth knowing
Without the licence the whole Cinematography department is absent from the project tree — not greyed out, not padlocked, simply not there.
Worth knowing
Training is real GPU work on hardware you provide. There is no hosted training tier.
Worth knowing
The rights to the footage or images you train on are yours to clear. The product records where a style came from; it does not tell you whether you were allowed to use it.
With the others
LookBook on its own is complete
Nothing below is required. Each line is what happens if you already own the other one.
+ VisualBook
The reason it exists. A published style is what a book is cast in, chosen once at Pre-Production ▸ Cinematography.
A long-form video storybook+ LitCaster
Independent of LitCaster. Neither needs the other.
A published podcast episodeQuestions
About LookBook specifically
Ready to try it?
The Core is free and does not need this product. Start there, then ask us about a licence.
- Do I need VisualBook to use LookBook?
- To build and publish a style, no. To render a book in one, yes — the per-book choice lives in VisualBook's Pre-Production tree, and rendering is what VisualBook does.
- What can I train on?
- Video you have, images you upload, or an archive query. Whether you are allowed to train on a given source is a question the product cannot answer for you, and it does not pretend to.
- What happens to published styles if the licence lapses?
- The Cinematography surfaces go. Ask before you let a licence go, because what remains selectable is exactly the kind of thing that should be in writing rather than in a FAQ — see Pricing.