PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ye Tian, Ling Yang, Jiongfan Yang, Anran Wang, Yu Tian, Jiani Zheng, Haochen Wang, Zhiyang Teng, Zhuochen Wang, Yinjie Wang, Yunhai Tong, Mengdi Wang, Xiangtai Li

arXiv:2511.09611v3cs.CV

TL;DR

Sequential thinking-aware image generation can degrade on complex tasks, and existing benchmarks do not evaluate alignment between generated reasoning and the final image. MMaDA-Parallel uses parallel multimodal diffusion and trajectory-level semantic rewards; it improves Output Alignment over Bagel on ParaBench.

  • Problem

    Sequential thinking-aware models can degrade image semantic fidelity, while existing benchmarks evaluate final images against prompts but not reasoning-to-image alignment.

    Source: §1 · pp. 1–2

  • Method

    MMaDA-Parallel uses discrete diffusion for parallel text–image generation with bidirectional attention at each denoising step, then applies ParaRL semantic rewards along the trajectory.

    Source: §1 · pp. 3–4

  • Results

    6.9% gain in Output Alignment over Bagel on ParaBench, with comparable performance on single-modality metrics.

    Source: §1 · p. 4

  • Takeaways & Limitations

    The approach improves cross-modal alignment and semantic consistency, offering a more robust paradigm for thinking-aware image synthesis.

    Source: §5 · p. 11

    MMaDA's relatively limited training data constrains its capabilities and makes it difficult to consistently surpass larger models such as Bagel.

    Source: J LIMITATIONS AND FUTURE WORK · p. 22

Abstract

from arXiv · show

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

1 INTRODUCTION

Sequential thinking-aware models can degrade on complex image tasks, and existing benchmarks do not assess whether generated reasoning aligns with the final image. MMaDA-Parallel combines parallel text–image diffusion with trajectory-level semantic rewards; on ParaBench it improves Output Alignment over Bagel.

1 INTRODUCTION

  • Sequential reasoning can shift from the main subject to minor details, causing the generated image to contradict the instruction and its own reasoning.This failure can reduce semantic fidelity on complex tasks.Source: §1 · p. 1
  • ParaBench evaluates text and image quality and their alignment, finding that Bagel's degradation occurs in categories with weakest reasoning–image alignment.Earlier benchmarks assess the final image against the initial prompt but not intermediate reasoning or its alignment with the output.Source: §1 · pp. 2–3
  • A VLM-generated reasoning trace turns image-editing and generation examples into instruction–reasoning–output quadruplets for supervised fine-tuning, and MMaDA-Parallel shows higher output consistency than sequential baselines.The source examples pair input images, edit instructions, and output images.Source: §1 · pp. 2–3
  • MMaDA-Parallel uses purely discrete diffusion with bidirectional attention between text and image at every denoising step to alleviate autoregressive error accumulation.The framework maintains cross-modal interaction throughout generation.Source: §1 · pp. 2–3
  • ParaRL assigns semantic rewards along the denoising trajectory, providing finer-grained alignment supervision than final-output-only optimization.The paper describes trajectory-level optimization as a more granular signal than traditional output-level supervision.Source: §1 · pp. 3–4
  • 6.9% gain in Output Alignment over Bagel on ParaBench establishes state-of-the-art performance among open-source models, with comparable single-modality metrics.The reported comparison is for the final framework.Source: §1 · p. 4

2 RELATED WORK

Related work explores interleaved text–image generation, reasoning before image synthesis, and feedback after generation. Most existing methods remain sequential autoregressive pipelines, while parallel generation with greater interaction between output modalities remains underexplored.

2 RELATED WORK

  • Prior approaches include interleaved generation in Chameleon (Team, 2024) and Mogao (Liao et al., 2025), and reasoning before synthesis in Image-CoT (Guo et al., 2025b) and GoT (Fang et al., 2025).Bagel (Deng et al., 2025a) extends reasoning to image generation and editing.Source: §2 · p. 4
  • OmniGen2 (Wu et al., 2025b) and IRG (Huang et al., 2025a) use reflective reasoning after image generation to refine visual outputs iteratively.Their feedback is multi-turn textual feedback.Source: §2 · p. 4
  • Most existing methods use sequential autoregressive interleaving, which could limit direct cross-modal interaction and risk error accumulation from imperfect reasoning traces.The passage identifies parallel generation with more interaction between output modalities as an underexplored direction.Source: §2 · p. 4

3 MMADA-PARALLEL

The section identifies cases where reasoning degrades image synthesis and introduces ParaBench to assess reasoning, images, and their alignment. It then presents parallel multimodal diffusion and trajectory-level alignment rewards as the proposed approach.

3.1 FINDINGS AND BENCHMARKING ON THINKING-AWARE SYNTHESIS

  • About 23% of cases showed performance declines with thinking, especially complex compositional edits; degraded Bagel categories also had drops in reasoning quality and output alignment.The analysis links the declines to low-quality or vague reasoning that misguides image generation.Source: §3.1 · pp. 4–5
  • ParaBench evaluates 300 prompts—200 editing and 100 generation—across text, image, and cross-modal output alignment, addressing benchmarks that assess images alone.Its visual-output aspects include image quality, alignment, and consistency; text aspects include text quality and alignment.Source: §3.1 · pp. 4–5

3.2 BASIC ALGORITHM AND ARCHITECTURE

  • MMaDA-Parallel represents text and images as discrete tokens in one interleaved sequence, using bidirectional attention and a shared mask predictor for synchronous denoising.The same parallel framework predicts both modalities during training and jointly decodes them during sampling.Source: §3.2 · pp. 5–6, Figure 3 · p. 5
  • Training noises only output tokens and uses timestep-reweighted cross-entropy, with text weight 1/t and image weight 1 to stabilize image quality and output alignment.The input sequence remains static while output tokens are masked according to the sampled timestep.Source: §3.2 · p. 6
  • At each shared diffusion step, modality-specific schedulers determine how many text and image tokens to reveal, while bidirectional attention lets the modalities inform each other.Experiments use a linear text reveal schedule with semi-autoregressive confidence-based decoding and a cosine image reveal schedule with global confidence-based decoding.Source: §3.2 · p. 6

3.3 POST TRAINING WITH PARALLEL REINFORCEMENT LEARNING

  • Supervised fine-tuning adapts MMaDA using image-and-instruction examples augmented with VLM-generated reasoning traces, so reasoning and synthesis occur concurrently.The resulting training examples pair input image, instruction, reasoning trace, and output image.Source: §3.3 · p. 7
  • ParaRL rewards text-image alignment at each denoising step, but computes rewards and advantages only at selected steps to make trajectory optimization feasible.The method uses sampled steps rather than optimizing over the entire dense trajectory.Source: §3.3 · p. 7
  • Semantically meaningful intermediate text and image fragments allow direct alignment rewards without a dedicated process reward model; CLIP scores are normalized, clipped, and rescaled to [0,1].The score normalization uses CLIP-score statistics estimated on a random 1% subset of the training data.Source: §3.3 · p. 8

4 EXPERIMENTS

MMaDA-Parallel is evaluated on ParaBench against thinking-aware and image-only models, with analyses of parallel decoding and trajectory-level optimization. It achieves the highest Output Alignment among open-source models and shows gains from trajectory-level rewards and denser reward sampling.

4.1 IMPLEMENTATION DETAILS

  • The final model uses two-stage training: SFT on MMaDA-MixCoT followed by GRPO-based RL focused on the most challenging 10% of SFT examples.Evaluation uses GPT-4.1 as an LLM judge across six metrics covering text quality, image fidelity, and cross-modal alignment.Source: §4.1 · p. 8

4.2 MAIN RESULTS

  • MMaDA-Parallel produces more precise reasoning traces and stronger compositional results than Bagel in examples involving a melting cake, withered grass, and counting.Bagel's reasoning in these cases is described as vague or missing crucial details.Source: Figure 6 · p. 9, §4.2 · p. 10
  • Across editing and generation tasks, MMaDA-Parallel achieves the highest Output Alignment among open-source models and performs on par with Bagel in general text and image quality.It also narrows the alignment gap to GPT-4o and Gemini-2.5 while maintaining competitive text and image quality.Source: Table 2 · p. 10, §4.2 · p. 10

4.3 ANALYSIS OF KEY CONTRIBUTIONS

  • Parallel decoding substantially outperforms sequential MMaA-Sequential on key alignment metrics while maintaining comparable text and image quality.The sequential baseline generates text before images, with noise applied to one modality at a time during training.Source: Table 3 · p. 10, §4.3 · p. 11
  • Trajectory-level ParaRL gains text–image consistency and Output Alignment over output-level RL, with more stable training dynamics.The comparison aggregates rewards across denoising steps rather than computing them only on the final sample.Source: Table 4 · p. 10, §4.3 · p. 11, Figure 7 · p. 11
  • Using s = 3 or s = 4 substantially improves over s = 2; the final configuration uses s = 3 for a balance of performance and efficiency.The paper attributes more stable guidance to the denser reward signal.Source: Table 5 · p. 10, §4.3 · p. 11

5 CONCLUSION

The paper identifies degradation in sequential thinking-aware models on complex tasks and presents ParaBench analysis linking it to poor alignment between generated modalities. It proposes MMaDA-Parallel with trajectory-level ParaRL, and reports that the framework also surpasses Bagel when applied to larger-scale Lumina-DiMOO.

5 CONCLUSION

  • The authors conclude that MMaDA-Parallel with ParaRL improves cross-modal alignment and semantic consistency, offering a more robust paradigm for thinking-aware image synthesis.Their analysis with ParaBench finds a strong correlation between performance degradation in sequential thinking-aware models and poor alignment between generated modalities.Source: §5 · p. 11

A SCALING OF MMADA-PARALLEL

  • Applied to larger-scale Lumina-DiMOO, the parallel framework and ParaRL surpass Bagel in thinking-aware synthesis, supporting the method's scalability.Lumina-DiMOO shares a similar architecture with MMaDA and benefits from larger-scale data training.Source: A SCALING OF MMADA-PARALLEL · p. 12

B.1 QUALITATIVE RESULTS

The section points to additional qualitative results for thinking-aware image editing and generation.

B.1 QUALITATIVE RESULTS

  • Additional qualitative examples for thinking-aware image editing and generation are provided in Figures 10 and 11.Source: §B.1 · p. 12

B.2 QUANTITATIVE RESULTS

This block reports image-only benchmark comparisons and describes related work, discrete-diffusion preliminaries, diffusion-compatible policy optimization, and text/image sampling. The method details explain how parallel decoding is trained and performed.

B.2 QUANTITATIVE RESULTS

  • MMaDA-Parallel consistently improves image alignment over sequential MMaDA on RISEBench and GenEval, narrowing the performance gap with Bagel.The supplied passages report no numeric values for these comparisons.Source: §B.2 · pp. 12–13

C MORE RELATED WORK

  • Prior diffusion-language-model work had not explored parallel text–image co-generation, while most multimodal RL rewards focused on final outputs rather than the denoising trajectory.ParaRL instead uses stepwise semantic alignment during denoising.Source: C MORE RELATED WORK · p. 16

D.1 PRELIMINARIES OF DISCRETE DIFFUSION MODELS.

  • Discrete diffusion corrupts data through a fixed forward Markov chain and learns a reverse process to denoise it; absorbing-mask training reduces to cross-entropy prediction of corrupted tokens.The paper applies this mask-prediction paradigm to both text and image tokens.Source: §D.1 · pp. 16–17

D.2 GROUP RELATIVE POLICY OPTIMIZATION FOR DISCRETE DIFFUSION MODELS

  • GRPO likelihoods are difficult to compute for parallel diffusion, so the adapted objective estimates them using randomly masked inputs and supports stable policy optimization.The objective sums over masked tokens and uses masking-based approximations of policy likelihoods.Source: §D.2 · pp. 17–18

E SAMPLING DETAILS ON TEXT AND IMAGE

  • Text and image logits are predicted jointly, but use distinct confidence measures and masking schedules to suit their different generation requirements.Image decoding retains the highest-confidence tokens at each step; text is generated in left-to-right blocks with iterative non-autoregressive refinement within each block.Source: E SAMPLING DETAILS ON TEXT AND IMAGE · p. 18
  • In the described implementation, iterative refinement continues until all image tokens are finalized; generating a 512px image encoded into 1,024 discrete tokens takes 30 steps to decode.Source: E SAMPLING DETAILS ON TEXT AND IMAGE · p. 18

F DETAILS OF TRAINING DATASET CURATION

The training data pairs images and instructions with generated reasoning traces, then filters and enhances samples. ParaBench evaluates reasoning, image quality, and their cross-modal alignment on editing and generation prompts.

F DETAILS OF TRAINING DATASET CURATION

  • The curated training collection contains 150,000 samples: 100,000 editing pairs and 50,000 generation pairs, a 2:1 ratio.The dataset was created because public editing and generation data generally lacked intermediate reasoning traces.Source: F DETAILS OF TRAINING DATASET CURATION · p. 19
  • A multimodal model generates step-by-step rationales from each input/output image pair and instruction, converting source triplets into reasoning-augmented quadruplets.Qwen2.5-VL-7B is used to produce the reasoning text.Source: F DETAILS OF TRAINING DATASET CURATION · p. 20
  • The 150,000-sample collection is filtered for near-duplicates and low-quality images, with GPT-4o regenerating higher-fidelity images when instructions remain valuable.Qwen-VL scoring identifies samples with low-quality or visually unappealing images.Source: F DETAILS OF TRAINING DATASET CURATION · p. 20

G DETAILS OF PARABENCH

  • ParaBench contains 300 challenging prompts—200 for editing and 100 for generation—and evaluates intermediate reasoning alongside visual output.Editing prompts span five categories; generation prompts include creative scenes, multi-object compositions, and long descriptive narratives.Source: G DETAILS OF PARABENCH · pp. 20–21
  • Its six evaluation axes cover text quality and alignment, image quality and alignment, editing consistency, and reasoning–image Output Alignment.Output Alignment measures cross-modal consistency between generated reasoning and the final image.Source: G DETAILS OF PARABENCH · p. 21

H MORE IMPLEMENTATION DETAILS

Implementation details specify the model initialization, two-stage post-training regimen, trajectory sampling, and inference-time sampling settings.

H MORE IMPLEMENTATION DETAILS

  • Training starts from MMaDA-MixCoT weights with LLaDA-8B and MagVIT-v2, then uses 30,000 SFT steps followed by 10,000 ParaRL steps.ParaRL uses approximately 15,000 examples, or 10% of the SFT dataset.Source: H MORE IMPLEMENTATION DETAILS · p. 21
  • ParaRL samples three trajectory points, with the same selected steps within each rollout and uniform sampling across rollouts.The KL constraint is β = 0.0001.Source: H MORE IMPLEMENTATION DETAILS · p. 21
  • Inference predicts all text and image logits in one forward pass, using classifier-free guidance scales of 3.5 for images and 0 for text.Source: H MORE IMPLEMENTATION DETAILS · p. 21

I MORE ABLATION STUDIES

The ablations examine modality-specific loss weighting and compare fully parallel, semi-parallel, and sequential decoding strategies.

I MORE ABLATION STUDIES

  • Using wtext(t) = 1/t and wimg(t) = 1 stabilizes image training and gives the best overall performance; sharing schedules either destabilizes training or reduces alignment.Source: I MORE ABLATION STUDIES · p. 22, Table 9 · p. 22
  • Fully parallel decoding jointly generates text and image at every denoising step and achieves strong results without extensive textual priors.The authors suggest early image steps establish coarse layouts, while excessive initial text may bias attention toward irrelevant details.Source: I MORE ABLATION STUDIES · p. 22, Table 10 · p. 22
  • Semi-parallel decoding improves over fully sequential decoding by interleaving image generation with the latter half of text denoising.The sequential variant generates text first and uses it as the sole image-conditioning signal, leaving outputs vulnerable to cross-modal error propagation.Source: I MORE ABLATION STUDIES · p. 22, Table 10 · p. 22

J LIMITATIONS AND FUTURE WORK

The authors identify limited base-model data and modality-specific sampling and training as current constraints, while proposing broader generation scenarios for future work. The appendix provides prompts for six evaluation dimensions.

J LIMITATIONS AND FUTURE WORK

  • Limited training data constrains MMaDA’s fundamental capabilities, making it difficult to consistently surpass larger-scale models such as Bagel.Bagel benefits from substantially larger training corpora.Source: J LIMITATIONS AND FUTURE WORK · p. 22
  • Sampling and training strategies are not yet fully unified across modalities; the authors suggest more integrated interaction paradigms may further enhance performance.Source: J LIMITATIONS AND FUTURE WORK · p. 22
  • The authors plan to extend the paradigm to story generation and multimodal outputs combining text and images.Source: J LIMITATIONS AND FUTURE WORK · p. 22

K PROMPTS FOR EVALUATION

Related papers

Loading 2511.09611v3…