Dif4FF: Leveraging Multimodal Diffusion Models and Graph Neural Networks for Accurate New Fashion Product Performance Forecasting

TL;DR

NFPPF faces severe domain shifts when forecasting sales for entirely new fashion items. This paper introduces Dif4FF, a two-stage pipeline that first uses a multimodal score-based diffusion model conditioned on image, release date, and Google Trends to generate multiple sales trajectories, then refines these outputs with a two-graph GCN to produce a final forecast. The diffusion stage captures uncertainty and distributional structure for unseen items, while the GCN refinement aligns predictions with temporal and prediction-space relationships, yielding robust, accurate forecasts. On the VISUELLE dataset, Dif4FF achieves state-of-the-art results with improved MAE and WAPE and demonstrates resilience to domain shift, suggesting practical benefits for reducing overproduction and waste in fast fashion. Future work includes adding more data sources and moving toward an end-to-end system, potentially enhancing real-world NFPPF applications.

Abstract

In the fast-fashion industry, overproduction and unsold inventory create significant environmental problems. Precise sales forecasts for unreleased items could drastically improve the efficiency and profits of industries. However, predicting the success of entirely new styles is difficult due to the absence of past data and ever-changing trends. Specifically, currently used deterministic models struggle with domain shifts when encountering items outside their training data. The recently proposed diffusion models address this issue using a continuous-time diffusion process. Specifically, these models enable us to predict the sales of new items, mitigating the domain shift challenges encountered by deterministic models. As a result, this paper proposes Dif4FF, a novel two-stage pipeline for New Fashion Product Performance Forecasting (NFPPF) that leverages the power of diffusion models conditioned on multimodal data related to specific clothes. Dif4FF first utilizes a multimodal score-based diffusion model to forecast multiple sales trajectories for various garments over time. The forecasts are refined using a powerful Graph Convolutional Network (GCN) architecture. By leveraging the GCN's capability to capture long-range dependencies within both the temporal and spatial data and seeking the optimal solution between these two dimensions, Dif4FF offers the most accurate and efficient forecasting system available in the literature for predicting the sales of new items. We tested Dif4FF on VISUELLE, the de facto standard for NFPPF, achieving new state-of-the-art results.

Figures (3)

• Figure 1: Dif4FF: a two-stage pipeline for New Fashion Product Performance Forecasting (NFPPF). Starting from multimodal signals (i.e., the image, the release data, and the Google Trends) of a single fashion product, we build a multimodal score-based diffusion model to generate an initial prediction of the sales, addressing potential objects with features beyond the training distribution. Then, we refine the Diffusion output using a powerful Graph Convolutional Network (GCN) architecture to obtain the final prediction.
• Figure 2: An overview of our multimodal score-based diffusion model. Each block contains two outputs: one for the subsequent block and another for a skip connection. The summation of all skip connections forms the model's final output. The primary component of each block is typically an S4 block gu2021efficiently.
• Figure 3: In the figures above are presented some visual representations of the multimodal score-based diffusion model outputs. In particular, the blue region represents the output distribution of the diffusion model given a certain sample. Specifically, the blue area is obtained by computing the weekly quantiles among the 50 outputs. The Prediction line, on the other hand, is the output of the refinement GCN, i.e., the final prediction. The forecasting period is six weeks from the release date, depicted on the x-axis. On the y-axis, the number of units sold of a specific garment in the various shops is shown.