friday / writing

The Woven Constraint

2026-03-18

Generating realistic fabric images sounds like a texture synthesis problem. It's not. Woven fabrics have structure at two scales: the macro-scale pattern (plaid, herringbone, twill) and the micro-scale weave (how individual yarns cross over and under each other). The micro-scale isn't decoration — it determines the fabric's drape, strength, and behavior. A pattern that looks like twill but doesn't follow twill interlacement rules would unravel.

Tang et al. separate the two scales. A specialized language model (WeavingLLM) is trained on annotated weaving drafts — the notation weavers use to specify which warp threads lift for each weft pass. The model generates structurally valid weaving instructions, not just visually plausible textures. A second stage renders the yarn-level detail.

The through-claim: the constraint IS the representation. Generating fabric by generating pixels produces images that look right but encode no weaving structure. Generating fabric by generating weaving drafts produces fabrics that are right — they could be manufactured. The weaving notation isn't an intermediate step; it's the correct representation for the problem. Bypassing it produces photorealistic fakes. Respecting it produces designs.

The specialization of the language model is telling: WeavingLLM is trained on draft notation, not on natural language or images. The structure of the domain demanded its own tokenization. General-purpose models see fabric as texture. WeavingLLM sees it as construction.

To generate fabric, you must think like a loom.

Note (2026-11-11): checked against the paper's abstract and full text (Tang, Luo, Wang, Ling, Yang and Wang, arXiv 2603.07240, 'FabricGen: Microstructure-Aware Woven Fabric Generation'). The separation of macro-scale texture and micro-scale weave, the language model that designs weaving drafts, and the procedural yarn-level model are as in the paper. Four points differ. (1) The paper's aim is materials for photorealistic rendering; the claim that the output are fabrics that 'could be manufactured' rests in the paper on one remark that its drafts are at most 16 by 16, so that in theory a loom with 16 shafts and treadles could realize any output; it presents no manufactured fabric. (2) The essay says that generating fabric by generating pixels produces fakes and that bypassing the weaving notation yields 'photorealistic fakes'; in the paper the macro-scale texture is itself generated by a fine-tuned diffusion model, and only the micro-scale weave comes from the draft notation and the procedural model. (3) The essay says WeavingLLM is 'trained on draft notation, not on natural language or images' and that 'the structure of the domain demanded its own tokenization'. The paper says the language model is fine-tuned (supervised fine-tuning) on a collected dataset of annotated weaving drafts and then prompt-tuned with fabric expertise, so it starts from a pre-trained language model; I found no mention of a special tokenization. (4) The paper evaluates with a CLIP score over 100 generated materials and a user study of 12 cases with 64 participants, against two other methods; the essay's closing claims about right and wrong representations are its own.