Paper 2604.20796
LLaDA2.0-Uni: Unifying Multimodal Understanding and Generation with Diffusion Large Language Model
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 5
- GitHub
- 772 stars
01 In brief
Summary
LLaDA2.0-Uni is a unified discrete diffusion large language model (dLLM) that integrates multimodal understanding and generation in a single framework.
Its architecture combines a SigLIP-VQ semantic tokenizer, a 16B MoE dLLM backbone, and a diffusion decoder.
The tokenizer converts images into discrete semantic tokens, enabling block-level masked diffusion for both text and vision.
The decoder reconstructs high-fidelity images from these tokens, with few-step distillation for efficient inference.
The model supports interleaved generation and reasoning, and achieves competitive performance on benchmarks like MMStar (64.1), MMMU (50.1), GenEval (0.89), and DPG (87.76), matching specialized VLMs and unified models.
Training uses a multi-stage pipeline with data packing and pre-extracted image tokens.
SPRINT, a training-free acceleration method, provides up to 1.6x speedup with minimal quality loss.
The diffusion decoder achieves 11.4x speedup via distillation (8 steps vs 50) with negligible performance drop.
Future work includes enhancing visual detail, scaling interleaved capabilities, and refining reinforcement learning.
02 From the paper
Abstract
We present LLaDA2.0-Uni, a unified discrete diffusion large language model (dLLM) that supports multimodal understanding and generation within a natively integrated framework. Its architecture combines a fully semantic discrete tokenizer, a MoE-based dLLM backbone, and a diffusion decoder. By discretizing continuous visual inputs via SigLIP-VQ, the model enables block-level masked diffusion for both text and vision inputs within the backbone, while the decoder reconstructs visual tokens into high-fidelity images. Inference efficiency is enhanced beyond parallel decoding through prefix-aware optimizations in the backbone and few-step distillation in the decoder. Supported by carefully curated large-scale data and a tailored multi-stage training pipeline, LLaDA2.0-Uni matches specialized VLMs in multimodal understanding while delivering strong performance in image generation and editing. Its native support for interleaved generation and reasoning establishes a promising and scalable paradigm for next-generation unified foundation models. Codes and models are available at https://github.com/inclusionAI/LLaDA2.0-Uni.