LLaDA2.0-Uni reasons before it draws, in a single diffusion MoE
- Source
- x.com
- Date
👋 Introducing LLaDA2.0-Uni — the 💥first unified MoE multimodal model in the LLaDA2.0 series of @TheInclusionAI, designed for native multimodal understanding and generation. #dLLM #inclusionAI #LLaDA #Multimodal #Opensource #DiffusionModels 1/2 Key features: 🧠Chain-of-Thought Generation LLaDA2.0-Uni doesn't just generate images — it thinks first. Through reasoning-augmented training, the model performs step-by-step reasoning before visual generation. Achieves 0.78 on WISE-Bench with thinking mode.
2/2 Key features: 🔄Interleaved Generation & Reasoning Unified discrete representations enable seamless interleaved text-image sequences. From story telling to science problems with step-by-step visuals — all in one coherent generation flow.
🔬 Built on three core model designs: 1️⃣ LLaDA2.0 Backbone for Unified Discrete Modelling: Leveraging the MoE dLLM architecture from LLaDA2.0, it formulates both multimodal understanding and generation as a unified block-wise mask prediction paradigm. 2️⃣ Fully Semantic Visual Tokens: Departing from traditional VQ methods that focus on reconstruction, LLaDA2.0-Uni transforms image inputs into purely semantic discrete tokens. 3️⃣ Diffusion Decoder for High-Quality Generation: Leveraging a custom Diffusion Decoder optimized for semantic tokens and few-step distillation, LLaDA2.0-Uni delivers high-fidelity image generation in just 8 steps of ultra-fast inference.
🔧Try out and explore how #LLaDA understanding and generate the world. 📄 Technical Report: https://t.co/DxlLxJ3zMy 🐙 Code: https://t.co/sX0lALW2AA 🤗 Model: https://t.co/n15P5dO7jS
Context
Most systems that both understand and generate images chain together two separate models: one that reasons about text and one that draws. Ant Open Source's LLaDA2.0-Uni, built under InclusionAI on the backbone of the company's earlier LLaDA2.0 diffusion release, instead handles both in one model: it reports performing reasoning before generating an image, and interleaving text and images within a single generation pass. The model converts images into purely semantic discrete tokens rather than the reconstruction-focused tokens most vision models use, then decodes them through a diffusion decoder the team says needs only eight steps.
Ant reports a score of 0.78 on WISE-Bench in the model's reasoning mode. The two demo clips in the announcement are animated GIFs with no accessible source file, only preview thumbnails, so this explanation is limited to the architecture and the one score the team states rather than the visual quality shown in those clips.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
- chain-of-thought — Having a model write out intermediate reasoning steps before its answer, which markedly improves performance on hard problems.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
Checking sign-in…
Loading comments…