Clothing pattern generation gets a two-way leash: DSG-BDF couples a design semantic graph with ControlNet structure, reaching FID 19.8
Synopsis
Controllable clothing-theme pattern generation has relied on text prompts or a separate structural branch, leaving the semantic and structural conditioning paths uncoupled during denoising and prone to semantic drift and style fluctuation. DSG-BDF encodes design-theme text with CLIP ViT-L/14 into a design semantic graph, turns it into structured semantic vectors through a graph neural network, and lets that path and a ControlNet contour, symmetry and layout path modulate each other during denoising. Under a unified setting it reaches CLIP-Score 0.347±0.003, FID 19.8±0.4 and SCI 0.912±0.012, above the strongest baseline FashionDiff (CLIP-Score 0.341, FID 20.9).
Interpretation
Clothing-theme pattern generation gains semantic and structural conditioning paths that adjust each other during denoising, rather than two condition types stacked side by side. In earlier controllable diffusion methods ControlNet mainly handled structural control and CLIP mainly handled semantic alignment, as independent paths; FashionDiff kept its structural modulation weights static throughout denoising, and PatternNet's structural and semantic pathways had no bidirectional interaction. Under unified prompts and random initialization, DSG-BDF reaches CLIP-Score 0.347±0.003, FID 19.8±0.4 and R-Precision 0.742±0.006, against the strongest baseline FashionDiff at CLIP-Score 0.341±0.003, FID 20.9±0.5 and R-Precision 0.728±0.006, with a paired two-tailed t-test (n=5) giving p=0.003.
The semantic condition is carried by a graph describing relationships between pattern elements and style attributes, not by a single text vector. A single vector embedding struggles to express combination relationships and style dependency structure among pattern elements, producing semantic ambiguity under complex theme descriptions. Design-theme text is encoded by CLIP ViT-L/14, tokens are split into element and style nodes by Stanford CoreNLP dependency labels, edge weights combine the text self-attention matrix with node feature similarity, and graph convolution plus global aggregation yields the structured semantic vector.
The structural condition is split into contour, symmetry-axis and layout-density channels and injected into the denoising backbone through a ControlNet branch. Text semantic conditions alone struggle to hold pattern structure stable, while contour, symmetry relations and layout rules are spatial constraints. Across eight geometric organizations (symmetric, repeating, radial, grid, striped, checkerboard, scattered motif, free layout), DSG-BDF's SCI is 0.927 for symmetric, 0.915 for grid, 0.908 for repeating, 0.903 for striped, 0.902 for radial, 0.898 for scattered motif and 0.892 for free layout, with standard deviations between 0.009 and 0.026.
The bidirectional collaborative module mainly resolves conflicts between the two conditioning paths during denoising rather than directly amplifying semantic consistency. Configuration M6, which introduces DSG and ControlNet together without collaborative modulation, reaches CLIP-Score 0.340, SCI 0.901 and style stability 0.823. The full model improves over M6 by 0.007 in CLIP-Score, 0.011 in SCI and 0.061 in style stability; without fine-tuning on Fashionpedia and ModaNet, the M6-to-M7 style stability gain is 0.061 and 0.059 respectively.
Perspective
The demonstrated scope is clothing-theme pattern generation at 512×512 resolution, with training data drawn from a DeepFashion subset, a pattern-classification dataset subset, manually curated pattern images and augmented semantic samples; cross-dataset testing covers only Fashionpedia and ModaNet without further fine-tuning. Inference runs on a single NVIDIA A100 80GB with 50 sampling steps, averaging 2.03 seconds per image at 11.4 GB peak memory and 1252M parameters. A reader concerned with open-vocabulary design semantics or higher resolution should watch for later work.
CLIP ViT-L/14 serves both as a training supervision signal and as the evaluation encoder, a metric-training circularity the authors partly mitigate with DINO-Similarity and FID. The perceptual study uses 30 evaluators with fashion or visual design backgrounds, a panel size that limits the power to separate closely ranked models. Semantic graph modeling relies on discrete semantic units and relational structure descriptions, so complex visual styles with high-level design context face representational compression in the graph. Of the training annotations, 80% come from rule-based dependency parsing and 20% from three annotators, and inconsistent style-attribute labeling introduces noise into semantic graph construction.
