AIAny
Icon for item

LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

Builds high-fidelity image generators by pairing a 6B Diffusion Transformer with a frozen LLaDA2.0-Mini vision-language module, relying on extensive image-only pretraining and mid-training; model weights, training code, and recipes are released.

Introduction

Why this matters Most open image generators are assembled from mixed recipes where visual priors and multimodal alignment are trained together; this paper separates those concerns. LLaDA-Image first builds a strong visual generative prior via image-only pretraining and mid-training, then pairs that prior with a frozen diffusion-language understanding module to produce photorealistic images that better follow fine-grained editing instructions.

Key Findings
  • Training recipe: uses a 6B Diffusion Transformer (DiT) trained from scratch with parameter-free RMSNorm and the Muon optimizer to improve scalability and stability. This design emphasizes optimizer and normalization choices as part of the recipe rather than only model scale.
  • Data strategy: reports a generation pipeline of 220M samples (paper text: "98 of which are real images"), highlighting a heavy reliance on image-only data before introducing multimodal alignment.
  • Distillation and speed: distills the model into LLaDA-Image-Turbo to enable fast inference in 2–4 sampling steps, trading some generative fidelity for latency.
  • Evaluation: achieves state-of-the-art open-source scores on Qwen-Image-Bench (53.53 English, 53.38 Chinese), indicating competitive alignment between image quality and instruction following.
  • Open research posture: authors release model weights, training code, and detailed recipes to facilitate reproducibility and further experimentation.
Who it's for and tradeoffs

Great fit if you are a research engineer or team that needs an open, end-to-end training recipe for high-quality diffusion-based image generation and instruction-guided editing — especially if you plan to reproduce, modify, or build on a clearly documented training pipeline. Look elsewhere if you need turnkey commercial services or models tuned for strict privacy/compliance constraints without re-training: the approach assumes substantial compute and large image datasets, and replicating results requires significant resources. The paper also embeds tradeoffs between dataset scale, compute, and distillation steps (Turbo) when prioritizing latency versus maximal image fidelity.

Where it fits

Positioned between research-focused diffusion model papers (method/recipe disclosure) and multimodal systems that jointly train vision-language alignment. Its main contribution is empirical: showing that a separately trained visual generative prior plus a frozen multimodal understanding backbone can yield both high visual quality and instruction-following capabilities while keeping training recipes fully open.

Information

  • Websitearxiv.org
  • OrganizationsAGI Research Center, Inclusion AI
  • AuthorsChuyan Chen, Haoxing Chen, Kun Chen, Zhenglin Cheng, Long Cui, Ruishan Fang, Zhangxuan Gu, Zhicheng Huang, Zhenzhong Lan, Yuanting Lei
  • Published date2026/09/03

More Items

Learns generalizable World Action Models for robotic manipulation by scaling causal egocentric video pretraining and grounding learned dynamics with heterogeneous robot trajectories. Key features: a three-stage curriculum (video pretraining, video-action mid-training with a unified action representation, and target-robot specialization) and a Slow–Fast dual-system for 30 Hz real-time action prediction.

Develops a vision-language foundation model for autonomous driving that unifies 3D BEV perception, visual question answering, and motion planning without changing the pretrained VLM architecture. Key elements include an external BEV perception head for 3D detection and occupancy, a Planning Expert using flow-matching for trajectory prediction, and a staged training recipe combining driving and general VLM data.

Converts posed indoor RGB(-D) video into editable, simulation-ready 3D scene graphs by parsing multi-view evidence into per-object bundles, generating complete object assets from that evidence, and placing them with GizmoAct, a VLM policy that refines 9-DoF poses through closed-loop GUI actions.