GARMENT
SKETCH
A SKETCH-TO-FASHION
BENCHMARK

Fashion sketching lets designers visualise a concept long before any fabric is cut, yet sketch-based fashion image synthesis has stalled for want of large-scale, high-quality paired data. GarmentSketch closes that gap: 26,249 fashion sketches across 21 garment categories, each paired with a detailed textual description. Captions were produced through a multi-stage pipeline combining several multimodal language models with human-in-the-loop refinement, balancing semantic accuracy against descriptive richness. Benchmarking state-of-the-art generators on the set exposes both the promise and the present limits of sketch-guided text-to-image generation and a clean trade-off between photorealism and faithfulness to the drawn line.

A sketch is sparse: a few abstract lines, no texture, no colour, no material cues. A garment photograph is dense with exactly those signals. Generators trained on general-purpose data struggle to span this gap they lose fabric draping, silhouette, and decorative detail, and lack the fashion-specific knowledge a professional workflow needs.

The central hypothesis: aligning rich textual semantics with sparse sketches bridges the modality gap, letting a model preserve structure and synthesise complex detail at the same time. That requires paired data which did not previously exist at scale.

Prior fashion datasets advance recognition, retrieval, and virtual try-on, but omit the sketch-plus-caption pairings needed for early-stage design. General sketch datasets, in turn, are either too simplistic or aimed at generic object retrieval rather than fine-grained garment structure.

Two parallel workflows turn a source image into a sketch–caption pair. Sketches come from an anime-style Informative-Drawings model about 6 seconds per image, versus roughly 10 minutes for stroke-optimisation tools, and far better at capturing intricate garment detail. Captions are synthesised by three multimodal models, then consolidated and human-verified.

Sources span Western e-commerce imagery, an upperwear top-up, and 650 hand-collected images of Eastern traditional dress such as the Vietnamese áo dài for cultural balance. The authors are candid about two biases: the set skews toward accessories, and remains Western-centric overall.

Shoes (27.9%) and bags (11.6%) together make up nearly 40% of the data, while core apparel like upperwear (8.06%) and bottomwear (10.2%) is thinner. The skew means models may learn the fixed structures of accessories more readily than clothing a limitation the authors flag as motivation for more globally balanced curation.

Four sketch-to-image models were evaluated zero-shot on the test set: Gemini 2.5 Nano Banana, ControlNet Scribble SDXL, ControlNet Scribble SD1.5, and T2I-Adapter Sketch SDXL. FID measures quality and diversity, LPIPS perceptual similarity to ground truth, CLIPScore semantic agreement with the prompt.

Per category, Gemini takes best FID in 17 of 21 classes and best LPIPS in 20 of 21. CLIPScore is more balanced: ControlNet Scribble SDXL leads 8 categories, Gemini 7, T2I-Adapter 4. SD1.5 trails throughout and stumbles hardest on culturally specific garments like the áo dài, reflecting limited diversity in its training data.

No model wins on every axis. The benchmark surfaces one clean tension: the more photorealistic a model's output, the more it tends to drift from the drawn structure and vice versa. Each model lands somewhere along this line.