Structured garment research predates the recent agent wave

DressCode joined natural-language guidance to two distinct outputs: SewingGPT generates sewing patterns with text guidance, while a separate Stable Diffusion branch generates tile-based physically based rendering textures.[S04] Its stated target is CG-friendly garments through natural-language interaction, with pattern completion and texture editing supporting the design process.[S04] That combination matters because a sewing representation and a surface appearance can be produced within one research framework without becoming the same artefact. A texture can describe the rendered surface while leaving panel dimensions, seam allowances and production tolerances unresolved.

The paper's production language therefore needs an artefact boundary. A text prompt expresses intent. A generated sewing representation encodes candidate construction information. A rendered result presents an appearance for review.[S04][S09] The three can be linked in a workflow, but approval of one does not automatically approve the others. A designer may accept the rendered collar proportion while a technician rejects the corresponding panel relationship. Conversely, a structurally valid pattern may produce an appearance that the designer did not intend. The research output becomes useful only when the review route preserves those separate decisions.

Garment3DGen illustrates a different route. It starts from a base mesh and an input image, uses image-to-three-dimensional generation as pseudo ground truth, then deforms the template mesh while preserving mesh quality and topology for draping and simulation.[S05] This is not DressCode's explicit text-to-sewing-pattern route. A simulatable mesh can be valuable for visualisation without supplying an editable flat pattern or a factory tech pack.[S04][S05][S22] The word garment in both titles does not make their outputs interchangeable.

Three input-output cards make the distinction operational. DressCode receives text and emits sewing patterns plus rendering textures, leaving factory specification outside its demonstrated CG-friendly objective.[S04] Garment3DGen receives an image and base mesh and emits a textured simulation asset, leaving explicit sewing-pattern export outside the abstract's claim.[S05] Sewformer receives a daily photograph and predicts a sewing pattern after training on synthetic pairs, leaving the transfer from its evaluated setting to a maker's product and material records unresolved.[S03] Each card identifies a research contribution and a boundary; none supplies a comparative production score.

This is why the chapter does not rank the methods by a single maturity number. A common score would require a common task, denominator and acceptance rule. The opened abstracts do not provide that shared protocol across text-guided pattern generation, template-mesh deformation and photo-based reconstruction.[S03][S04][S05] GFT can legitimately ask which artefact is needed next, then test the candidate for that artefact. It cannot infer factory readiness from the richness of a rendered result.

Later research also exposes variables that an attractive garment image leaves unstated. Image2Garment addresses the absence of material properties in geometry-only prediction by inferring material composition and fabric attributes, then mapping those attributes to physical fabric parameters through a predictor trained on a small dataset of material-physics measurements.[S11] FIT addresses a different omission: appearance-driven virtual try-on can depict a garment as well fitted without adequately representing the person's and garment's sizes.[S13] EASE addresses yet another: local material allowance is often a byproduct of geometry or simulation rather than an independent design variable.[S27] These contributions should not be collapsed into a general claim that models now understand garments.

An input-output card keeps the first boundary clear. Image2Garment receives a single image and produces simulation-ready garment estimates that include inferred material information and predicted physical parameters.[S11] Its research contribution is the connection between visual attributes and simulation inputs. Its limitation for a maker is equally specific: an inferred composition or parameter is not a measurement certificate for the fabric actually purchased. The small measurement dataset grounds the research method, but does not turn every subsequent image-based estimate into a local material observation.[S11]

FIT receives the virtual try-on problem as its research boundary rather than treating it as solved by image realism. The authors introduce synthetic try-on triplets accompanied by body and garment measurements, including ill-fit cases created through programmatic garments and physics simulation.[S13] The output is a dataset and a fit-aware baseline model, not a physical fit approval for a wearer. Its importance lies in evaluation design: deliberately mismatched sizes make a previously hidden failure visible. It does not follow that every commercial try-on system has the same failure or that the new baseline resolves comfort, movement or buyer acceptance.[S13]

EASE begins with a parametric body surface and represents ease as spatially varying, anisotropic per-triangle scales. It then optimises sewing patterns intended to preserve the specified ease while satisfying geometric and seam constraints.[S27] This makes ease an editable input rather than merely an outcome observed after simulation. The capability does not establish which ease distribution a Cameroonian wearer population prefers, nor whether a chosen fabric and activity will feel comfortable. Those remain product and physical-validation questions rather than transferable accuracy claims.[S27]

The three cards therefore answer different questions. Image2Garment asks how an image might inform material and simulation parameters. FIT asks whether size information and ill-fit examples enter try-on evaluation. EASE asks how a designer can specify and transfer local allowance.[S11][S13][S27] A combined accuracy league table would erase these distinctions. It would also hide the difference between inferred material properties, synthetic fit data and an explicit design representation.

For GFT, the chronology supports a more modest inference: research has made missing variables easier to name and test. A local evaluation should carry each variable into its own evidence field, with the source of the value, the evaluated case and the unresolved physical check recorded separately.[S20] Material estimation can propose a swatch test, fit-aware data can inspire matched and mismatched cases, and explicit ease can inform a specification review. None of these research results replaces the corresponding local evidence.