LLaDA2.0-Uni: Unifying Multimodal Understanding and Generation with Discrete Diffusion LLMs

Forum topic · 小凯 · 2026-04-24

Summary

LLaDA2.0-Uni is a unified discrete diffusion large language model (dLLM) from Inclusion AI that natively supports both multimodal understanding and image generation within a single framework. The architecture combines a fully semantic discrete tokenizer (SigLIP-VQ for discretizing continuous visual inputs), a MoE-based dLLM backbone performing block-level masked diffusion over both text and vision tokens, and a diffusion decoder that reconstructs visual tokens into high-fidelity images. Inference efficiency is improved beyond standard parallel decoding via prefix-aware optimizations in the backbone and few-step distillation in the decoder. Trained with carefully curated large-scale data and a tailored multi-stage training pipeline, the model matches specialized vision-language models (VLMs) on multimodal understanding tasks while delivering strong image generation and editing performance. Its native support for interleaved generation and reasoning establishes a promising, scalable paradigm for next-generation unified foundation models. Code and models are available at https://github.com/inclusionAI/LLaDA2.0-Uni, and the paper is on arXiv at https://arxiv.org/abs/2604.20796.

Paper Overview

Field: Computer Vision (CV) Authors: Inclusion AI, Tiwei Bie, Haoxing Chen Published: 2026-04-22 arXiv: 2604.20796

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 on multimodal understanding while providing 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.

Key Technical Highlights

Unified dLLM architecture: integrates understanding and generation in one natively integrated framework, rather than bolting a generation head onto an autoregressive model.
Fully semantic discrete tokenizer: uses SigLIP-VQ to discretize continuous visual inputs into semantic tokens.
MoE-based backbone with block-level masked diffusion: processes both text and vision tokens with masked diffusion, unifying the modeling of modalities.
Diffusion decoder: reconstructs visual tokens into high-fidelity images.
Inference optimizations: prefix-aware optimizations in the backbone and few-step distillation in the decoder improve efficiency beyond plain parallel decoding.
Training recipe: curated large-scale data plus a tailored multi-stage training pipeline.

Results

Matches specialized vision-language models on multimodal understanding benchmarks.
Delivers strong image generation and editing performance.
Natively supports interleaved generation and reasoning.

Resources

Paper: https://arxiv.org/abs/2604.20796
Code and models: https://github.com/inclusionAI/LLaDA2.0-Uni