ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion

2026-04-10

AI Summary

This work addresses the inefficiency of existing approaches in chest X-ray report generation, where conventional autoregressive models suffer from high inference latency and diffusion-based methods either require slow multi-step denoising or sacrifice textual coherence in single-step generation due to mean-field bias. To overcome these limitations, the authors propose ECHO, an efficient diffusion-based vision-language model that accelerates inference through a block-wise one-step generation strategy. ECHO introduces a Direct Conditional Distillation (DCD) framework that leverages intra-policy diffusion trajectories to construct non-factorized supervision signals, effectively capturing joint token dependencies. Additionally, a Response-Asymmetric Diffusion (RAD) training strategy is devised to enhance training efficiency. Experiments demonstrate that ECHO achieves 64.33% and 60.58% relative improvements on RaTE and SemScore metrics, respectively, while offering an 8× faster inference speed without compromising clinical accuracy.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

Chest X-ray report generation (CXR-RG) has the potential to substantially alleviate radiologists'workload. However, conventional autoregressive vision--language models (VLMs) suffer from high inference latency due to sequential token decoding. Diffusion-based models offer a promising alternative through parallel generation, but they still require multiple denoising iterations. Compressing multi-step denoising to a single step could further reduce latency, but often degrades textual coherence due to the mean-field bias introduced by token-factorized denoisers. To address this challenge, we propose \textbf{ECHO}, an efficient diffusion-based VLM (dVLM) for chest X-ray report generation. ECHO enables stable one-step-per-block inference via a novel Direct Conditional Distillation (DCD) framework, which mitigates the mean-field limitation by constructing unfactorized supervision from on-policy diffusion trajectories to encode joint token dependencies. In addition, we introduce a Response-Asymmetric Diffusion (RAD) training strategy that further improves training efficiency while maintaining model effectiveness. Extensive experiments demonstrate that ECHO surpasses state-of-the-art autoregressive methods, improving RaTE and SemScore by \textbf{64.33\%} and \textbf{60.58\%} respectively, while achieving an \textbf{$8\times$} inference speedup without compromising clinical accuracy.

Problem

• Chest X-ray report generation
• diffusion models
• inference latency
• textual coherence
• one-step generation

Innovation

• one-step diffusion
• Direct Conditional Distillation
• Response-Asymmetric Diffusion
• chest X-ray report generation
• vision-language model