MMaDA-Parallel: Multimodal Large Diffusion Language Models for Thinking-Aware Editing and Generation

TL;DR

The paper identifies error propagation in autoregressive thinking-aware generation and introduces ParaBench to quantify text–image reasoning–output alignment. It then proposes MMaDA-Parallel, a parallel, discrete diffusion framework with interleaved text and image tokens and bidirectional cross-modal attention, trained first with supervised fine-tuning and then enhanced via Parallel Reinforcement Learning (ParaRL) that provides trajectory-level semantic rewards along the denoising path. Empirical results on ParaBench show state-of-the-art Output Alignment among open-source models and competitive unimodal quality, with a notable $6.9$ percentage-point improvement over Bagel; ParaRL further boosts cross-modal grounding and consistency. The work demonstrates scalability through experiments with Lumina-DiMOO and highlights the significance of trajectory-level supervision for robust thinking-aware multimodal generation, offering a practical paradigm for joint reasoning and editing across modalities.

Abstract

While thinking-aware generation aims to improve performance on complex tasks, we identify a critical failure mode where existing sequential, autoregressive approaches can paradoxically degrade performance due to error propagation. To systematically analyze this issue, we propose ParaBench, a new benchmark designed to evaluate both text and image output modalities. Our analysis using ParaBench reveals that this performance degradation is strongly correlated with poor alignment between the generated reasoning and the final image. To resolve this, we propose a parallel multimodal diffusion framework, MMaDA-Parallel, that enables continuous, bidirectional interaction between text and images throughout the entire denoising trajectory. MMaDA-Parallel is trained with supervised finetuning and then further optimized by Parallel Reinforcement Learning (ParaRL), a novel strategy that applies semantic rewards along the trajectory to enforce cross-modal consistency. Experiments validate that our model significantly improves cross-modal alignment and semantic consistency, achieving a 6.9\% improvement in Output Alignment on ParaBench compared to the state-of-the-art model, Bagel, establishing a more robust paradigm for thinking-aware image synthesis. Our code is open-sourced at https://github.com/tyfeld/MMaDA-Parallel

Figures (19)

• Figure 1: Sequential vs. parallel thinking-aware image synthesis. (a) Sequential generation (Bagel, GPT4o) may suffer from vague or incorrect reasoning. (b) Parallel generation aligns text and image at each denoising step, reducing hallucination and errors. (c) Quantitative comparison shows reasoning can degrade performance in certain categories. (d) Poorer categories also exhibit weaker reasoning–image alignment, highlighting the need for stronger cross-modal alignment.
• Figure 2: MMaDA-Parallel supports parallel, thinking-aware image editing and generation. Compared with Bagel, MMaDA-Parallel demonstrates superior reasoning quality and stronger alignment between the generated text and image outputs.
• Figure 3: Parallel Generation Architecture: During (a) training, image and text responses are masked and predicted in parallel with a uniform mask predictor, optimized by the masked token likelihood objective. During (b) sampling, the model performs parallel decoding to generate both image and text responses jointly, enabling efficient multimodal response generation.
• Figure 4: Overview of our proposed Parallel Reinforcement Learning (ParaRL). Rather than optimization only to the final denoised outputs, ParaRL introduces reward signals along the entire denoising trajectory, reinforcing semantic alignment consistently throughout the generation process.
• Figure 5: Synergy of sampling. Given the prompt: "change the blue shirt to a vibrant rainbow color," the specific color decoding in text and image emerges at the same step.