Training-Free Consistency Pipeline for Fashion Repose

TL;DR

FashionRepose introduces a training-free, zero-shot pipeline for pose normalization of long-sleeve garments that preserves garment identity and branding during non-rigid edits. It deploys a multi-stage architecture that combines off-the-shelf models (e.g., RealisticVision, ControlNet OpenPose/Canny, Florence2, SAM2) with custom CV techniques (silhouette-based shape matching, sleeves-torso composition, and logo reinsertion) to achieve near real-time editing without retraining. Quantitative and qualitative evaluations on DressCode and VITON-HD demonstrate competitive metrics (LPIPS, PSNR, SSIM) and clear qualitative gains in pose consistency and identity preservation, including in logo-bearing garments. The work highlights practical impacts for e-commerce and design workflows, while noting limitations with complex textures and potential artifacts from composition, and points to texture reconstruction and mask refinement as future directions.

Abstract

Recent advancements in diffusion models have significantly broadened the possibilities for editing images of real-world objects. However, performing non-rigid transformations, such as changing the pose of objects or image-based conditioning, remains challenging. Maintaining object identity during these edits is difficult, and current methods often fall short of the precision needed for industrial applications, where consistency is critical. Additionally, fine-tuning diffusion models requires custom training data, which is not always accessible in real-world scenarios. This work introduces FashionRepose, a training-free pipeline for non-rigid pose editing specifically designed for the fashion industry. The approach integrates off-the-shelf models to adjust poses of long-sleeve garments, maintaining identity and branding attributes. FashionRepose uses a zero-shot approach to perform these edits in near real-time, eliminating the need for specialized training. consistent image editing. The solution holds potential for applications in the fashion industry and other fields demanding identity preservation in image editing.

Figures (21)

• Figure 1: Pose Normalization Task Overview.FashionRepose presents a training-free, consistency-aware method for normalizing the pose of long-sleeve garments, starting from an initial still-life configuration. Our pipeline enforces a standardized 45-degree arm-to-torso alignment, preserving garment identity, texture fidelity, and brand-specific attributes across complex, non-rigid transformations (\ref{['sec:pipeline_architecture']}). Combining state-of-the-art models with custom computer vision algorithms, this scalable solution enables precise garment pose editing.
• Figure 2: Proposed Pipeline Architecture. The figure depicts the architecture of FashionRepose, a pose-normalization pipeline tailored for fashion domain. The pipeline is able to edit, in less than a minute, the pose of long-sleeve garments maintaining consistency, and identity. The entire workflow is entirely based on pretrained off-the-shelf models fostering ease of adoption through its training-free capability. The stages contained in the yellow section of the workflow are considered pre-pipeline stages to easily integrate the pipeline in already existing industrial processes (see \ref{['sec:long_sleeves_detection']}).
• Figure 3: Coarse Garment Generation. Leveraging Realistic Vision v5.1, we generate a first coarse pose-normalized clothing starting from a still-life logo-suppressed garment. During this stage, the Stable Diffusion model is conditioned by different embeddings coming from both the IP-Adapter and ControlNets.
• Figure 4: Conditioned Unsampling. From the still-life and the pose-normalized images, we obtain the latents and the canny-processed images. The latents are fused together through a gradient-mask blending operation. Then, a two-pass guiding process (unsampling - sampling) adds and removes conditioned noise.
• Figure 5: Garment Parts-Composition. The figure illustrates a graphical representation of \ref{['alg:parts_composition_algorithm']}. Left: The points identified by the algorithm are over-imposed on the normalized-pose mask. Right: The green lines represent the discriminating boundaries between the sleeves and the torso regions of the image.