ASIA: adaptive single-diffusion-stage inference acceleration for diffusion-based virtual try-on

Virtual try-on synthesises an image of a person wearing a chosen garment. Diffusion models give the most realistic results, but remain too slow for interactive use. The leading method, TPD, runs its diffusion network twice: a first pass predicts a pair-specific mask locating where the garment belongs, and a second pass renders the garment into that region. Each pass takes about 100 denoising steps, so one image needs roughly 200, half of them producing only a binary shape. This first stage does not justify a diffusion model. We propose Adaptive Single-Diffusion-Stage Inference Acceleration (ASIA), which replaces diffusion-based mask generation with a lightweight convolutional network. Using DensePose and the flat garment mask as structural priors, ASIA predicts the pair-specific mask in one forward pass, reducing that stage from seconds to milliseconds. Try-on then needs a single diffusion stage, whose input block alone is finetuned to the new mask distribution. On VITON-HD, inference falls from 15.4 to 7.5 seconds per image, a 2.05-fold speedup. Under a controlled reproduction of TPD, paired structural (SSIM) and perceptual (LPIPS) similarity both improve while the Fréchet Inception Distance rises slightly (9.04 against 8.78). Removing an entire diffusion stage thus trades a small FID increase for a large speedup and improved paired fidelity.