Original title: Controllable flow fields for high-fidelity virtual try-on with diffusion models
Kun Li, Lai-Man Po, Wing-Yin Yu, Yu Xue +5 · September 2026
Virtual try-on programs let computers show how clothes look on a person. Current tools need huge amounts of computer memory and lose tiny fabric details. To fix this, researchers built an end-to-end framework based on diffusion models, which generate images step by step. It uses a tool named COFFLOW to map out exactly how target clothes match up with a person's body. The system uses this map to guide how it draws the new outfit. Another module protects untouched areas like faces, arms, and legs so they stay sharp. The team kept most of the main image maker frozen and only trained the mapping tool, cutting training costs. The result fits clothes and poses better, keeps textures sharper, and uses fewer computer resources.
Imagine you want to paste a sticker of a shirt onto a photo of a person, but the sticker must stretch and fold to match the person's pose. If you just slap it on, the sticker looks flat and fake. Now suppose you first trace the person's body outline and then carefully align the sticker's corners to that outline, so it bends naturally with the arms and torso. In this paper, the COFFLOW module does that tracing and alignment for the AI, ensuring the garment image fits the person's shape realistically.
Online shopping for clothes often fails because you cannot tell how a garment will actually look on your body. This research makes virtual try-on more realistic by preserving tiny details like fabric texture and fit, which could reduce returns and improve the shopping experience. It also uses less computing power, making it practical for real e-commerce platforms.
The abstract does not discuss failure cases, such as handling complex garments with many layers or extreme poses. It also does not mention how the method performs on diverse body types or in real-time applications. The evaluation is limited to two benchmarks, so generalizability to unconstrained settings is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Controllable flow fields for high-fidelity virtual try-on with diffusion models — Kun Li, Lai-Man Po, Wing-Yin Yu et al. — Multimedia Systems — doi:10.1007/s00530-026-02660-9
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.