From Denoising to Refining: A Corrective Framework for Vision-Language Diffusion Model

TL;DR

This paper tackles the train-inference discrepancy that causes error cascades in discrete vision-language diffusion models during parallel generation. It introduces ReDiff, a refining-enhanced diffusion framework with two stages: foundational revision training to handle synthetic errors and an online self-correction loop where drafts are revised by an expert to train the model on its own mistakes. Through extensive experiments on caption benchmarks, ReDiff achieves state-of-the-art performance among diffusion-based VLMs and shows markedly improved stability and factual accuracy in few-step, parallel generation. The work offers a practical path toward robust, efficient, and controllable vision-language diffusion systems by enabling iterative content refinement during generation.

Abstract

Discrete diffusion models have emerged as a promising direction for vision-language tasks, offering bidirectional context modeling and theoretical parallelization. However, their practical application is severely hindered by a train-inference discrepancy, which leads to catastrophic error cascades: initial token errors during parallel decoding pollute the generation context, triggering a chain reaction of compounding errors and leading to syntactic errors and semantic hallucinations. To address this fundamental challenge, we reframe the generation process from passive denoising to active refining. We introduce ReDiff, a refining-enhanced diffusion framework that teaches the model to identify and correct its own errors. Our approach features a two-stage training process: first, we instill a foundational revision capability by training the model to revise synthetic errors; second, we implement a novel online self-correction loop where the model is explicitly trained to revise its own flawed drafts by learning from an expert's corrections. This mistake-driven learning endows the model with the crucial ability to revisit and refine its already generated output, effectively breaking the error cascade. Extensive experiments demonstrate that ReDiff significantly improves the coherence and factual accuracy of generated content, enabling stable and efficient parallel generation far superior to traditional denoising methods. Our codes and models are available at https://rediff-hku.github.io/.

Figures (8)

• Figure 1: Comparison between standard vision-language diffusion models and our proposed refining-enhanced approach. (a) Mask-pred diffusion is trained for passive denoising (mask recovering under fixed context). An initial error, such as misidentifying the "bus" as a "trunk", triggers an error cascade. The model cannot correct this mistake and proceeds to hallucinate further details based on the flawed context (e.g., "Two men are drinking"), leading to a factually incorrect description. (b) Refining-enhanced diffusion introduces a paradigm of active refining, teaching the model not only to predict masked tokens but also to perform token refinement. Our ReDiff learns to self-correct through an online loop where its own "flawed drafts" are revised by an expert revisor. As a result, the model can identify and correct its initial mistakes (revising "trunk" to "bus", "Two men" to "A man"), breaking the error cascade and generating a factually grounded response. (c) Performance comparison between LLaDA-V DBLP:journals/corr/abs-2505-16933/llada-v and ReDiff under different inference speeds. "CLAIR" and "Coverage" are detailed caption metrics on CapMAS DBLP:journals/corr/abs-2412-15484/capmas, and "CAPTURE" is on DetailCaps-4870 dong2024benchmarking/capture. Our model delivers superior generation quality and achieves more stable results when using fewer inference steps.
• Figure 2: Overview of our proposed two-stage training framework for corrective refining. (a) We illustrate common failure modes in standard vision-language diffusion models, which are prone to generating syntactic errors (e.g., "Domin bus bus") and factual hallucinations (e.g., "a woman"). (b) In the foundational revision training stage, we instill a general corrective capability by training a base model (ReDiff-Base) to revise synthetic errors that are intentionally injected into ground-truth captions. (c) For the second stage, i.e., online self-correction learning, the model generates its own flawed "drafts". These drafts, containing the model's intrinsic errors, are then revised by an expert AI assistant. The resulting "draft-refined pairs" provide strong supervision, teaching our final model (ReDiff) to identify and correct its own characteristic mistakes, thus breaking the error cascade.
• Figure 3: Cases comparison between LLaDA-V and our ReDiff under 4 tokens/step inference speed. ReDiff demonstrates superior fluency and accuracy in its generated captions.
• Figure 4: Refinement process of ReDiff at different inference step. Red tokens indicate the errors produced during generation, while green tokens mean the corresponding refined results.
• Figure 5: Generation results of ReDiff with or w/o refinement during inference.