A bespoke identity system that knows what it sounds like out loud.
Identity Loom is the work file for a five-month commission with a 117-year-old print-and-letterpress publisher in north Yorkshire. The client wanted a generative identity system — logotypes, masthead variants, ornamentation, secondary type — that could be steered by an in-house design lead without producing the flat, somebody-else's-accent output that foundation-model identity work tends to carry. The system shipped in October 2025 and has produced thirty-eight authored marks since.
A publisher who already had a voice.
The client came to the studio with a problem that we have heard, with variations, eleven times: they had a strong house identity built up across a century of letterpress books and quarterly literary supplements, and they were watching their younger designers reach for foundation-model image generators when they needed a fresh masthead, a programme cover, or a one-off ornament. The results were technically competent and immediately legible as not theirs. The accent was wrong; the kerning was right.
The brief, in the survey week, came to one sentence: build us a system that sounds like us when we talk to it. The partners on the file were Wren Halberd (corridor lead) and Pita Selafyan (operations, kill switch, data licence). The studio was Eira Ashlin and Marisol Vegh in reserve for the structural questions.
What we were given to learn from.
The archive was four full filing cabinets of letterpress proofs from 1908 to the present, plus a digital archive (Adobe InDesign, mostly, with some older QuarkXPress files) of every cover, masthead and ornament the client had typeset since 1996. The total corpus was roughly 12 000 individual marks — some monograms, some full-titlepage compositions, some single ornamental rules. Of those, 4 200 had been hand-signed by the in-house compositor; the signature pattern is itself a feature of the house identity.
We digitised the cabinet contents in week two using a flatbed scanner the studio owns. We did not outsource the scanning; the partner on a file is the partner who handles the data, because the data is the brief and we do not want anyone else's eyes on it before ours.
Three small models that speak in the same accent.
The shipped system is three small generative networks, not one. We considered a single multi-modal model and we discarded it twice; multi-modal in this regime gave us prettier marketing screenshots and worse production behaviour, because the three corridors (mark, ornament, secondary type) have different constraints and the model that did all three did each of them with less conviction than the focused ones.
Loom-A · The mark model. A 340M-parameter diffusion model trained on the masthead-and-monogram sub-corpus. Conditioned on a short prose prompt (the in-house design lead types two sentences about what the mark is for) plus a continuous "weight" vector that captures the publisher's compositor-era preference for asymmetric counters and short ascenders.
Loom-B · The ornament model. A 160M-parameter VAE trained on the ornamental-rule and decorative-initial sub-corpus. Smaller because the corpus is smaller and because ornaments do not need conditional steering — they need a controllable amount of variation around a fixed identity vector.
Loom-C · The secondary-type generator. A glyph-by-glyph autoencoder for producing variant cuts of the client's eleven in-house faces. Outputs are exported as OpenType source files; the design lead reviews and signs each glyph before the cut is committed.
The interface the client actually uses.
None of the three models is the product. The product is the loom — a small desktop application that sits in front of the three models and presents the design lead with a single survey-room metaphor: the lead types a prose brief, draws a constraint sketch on a digitiser pad, and selects which of the three models is being woven. The loom produces six candidates, the lead pins two, and the lead iterates from there. Every artefact is signed by the lead at export.
The loom is a sketch tool, not a generator. We are explicit about this in the model card and in the training materials. The lead is the designer; the loom is the bench. Marks are authored, not summoned.
loom rewind --everything. It removes the three models, every generated output, every cached embedding, and every training checkpoint from the client's infrastructure. We built it in week one. It has been tested three times in production rehearsals, never in anger.
One workstation. Five months. Four signatures.
The loom and its three models run on a single workstation in the client's design studio — an A6000-class machine that the client already owned for video work. Total inference latency for a six-candidate mark generation is 4.2 seconds. The largest single file in the deployed system is 1.3 GB; the entire installation fits inside the client's nightly backup.
We shipped on the 12th of October 2025, six days late against the survey-room plan. The lateness was caused by Loom-C's OpenType export pipeline; we underestimated the work involved in producing OpenType source files that would pass the client's existing pre-press validation. We absorbed the overrun on a fixed-fee contract because the survey-room estimate was ours to defend.
Every shipped artefact — the three models, the loom binary, the model cards, the kill-switch agreement, the training-data manifest — was signed by Wren Halberd as the partner on the file and counter-signed by Pita Selafyan for operations.
Thirty-eight marks in the first six months.
Between October 2025 and April 2026 the in-house design lead produced thirty-eight authored marks using the loom. Of those, twenty-one were programme covers, eleven were masthead variants for the quarterly supplement, and six were ornamental sequences for limited-edition letterpress reprints. The design lead reports the loom as "a faster pencil"; we read every quarterly review they send us and we treat that line as the success criterion we were aiming at.
The pieces we did not ship are equally informative: a fourth model that would have generated full page layouts on the basis of editorial briefs, considered in week three of the survey and discarded by the end of week four. The client preferred to retain the editorial-judgement step as a fully human one. We agreed in the room; we did not build it; we are glad we did not.
If your house has an archive and an accent.
Identity Loom is the corridor for clients who already have a voice — an archive, a compositor, a hundred years of marks — and who are watching the accent leak. We do not take this corridor on for new houses with no archive; the work depends on the existence of a corpus we can train into. If you are a publisher, a foundry, a small museum or a long-running studio, the brief sheet is the fastest route in.