FlashClear: Ultra-Fast Image Content Removal via Efficient Step Distillation and Feature Caching

2026-05-09

AI Summary

This work addresses the high computational redundancy and slow inference of existing diffusion-based image inpainting methods, which denoise the entire image at every timestep. To overcome this limitation, the authors propose Region-Aware Distillation (RAD), a training framework for efficient few-step models, coupled with a training-agnostic Foreground-Priority Asymmetric Attention and feature Caching strategy (FPAC) that substantially reduces computational overhead. Evaluated on the OBER benchmark, the proposed approach achieves speedups of 8.26× over ObjectClear and 122× over OmniPaint, while preserving high visual fidelity and generation quality.

Abstract

Recently, diffusion-based object removal models have achieved impressive results in eliminating objects and their associated visual effects. However, they indiscriminately denoise all tokens across all timesteps, ignoring that removal usually involves small foreground regions. This strategy introduces substantial computational overhead and prolonged inference times. To overcome this computational burden, we propose a latent discriminator to implement Region-aware Adversarial Distillation (RAD), yielding a highly efficient few-step model named FlashClear. Furthermore, tailored to few-step diffusion models, we propose FPAC (Foreground-Prioritized Asymmetric Attention and Caching), a training-free acceleration strategy. Extensive experiments demonstrate that our framework provides massive acceleration while maintaining or exceeding the performance of our base model, ObjectClear. Notably, on the OBER benchmark, our FlashClear achieves up to 8.26$\times$ and 122$\times$ speedup over ObjectClear and OmniPaint, respectively, while maintaining high visual quality and fidelity.

Problem

• image content removal
• diffusion models
• computational overhead
• inference acceleration
• object removal

Innovation

• Region-aware Adversarial Distillation
• Few-step Diffusion
• Feature Caching
• Foreground-Prioritized Attention
• Image Inpainting Acceleration