Paper Walkthrough: LLaDA-Image — Building the Visual Prior from Images Alone, Then Distilling to 2–4 Steps

A training recipe that builds the visual prior from images alone before any caption enters the picture. We walk through LLaDA-Image — a 6B DiT trained from scratch and distilled down to 2–4 sampling steps — following the design decisions the paper actually makes.

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

Learning to Draw Before Learning to Listen

The paper is "LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes" (arXiv:2609.03796, published 3 September 2026, by the AGI Research Center at Inclusion AI).

Here is what it claims. LLaDA-Image is a unified framework pairing a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA 2.0 Mini diffusion language model backbone. Rather than leaning heavily on paired image–text data from the start, the authors first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98% of which are real images. For efficient and scalable optimization, they use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while following fine-grained editing instructions. They further distill it into LLaDA-Image Turbo for fast inference in 2–4 sampling steps. On Qwen-Image-Bench it scores 53.53 on the English track and 53.38 on the Chinese track, a new state of the art among open-source models on both. Weights, training code, and detailed recipes are released.

What makes this report worth reading is the procedure, not the leaderboard row. The paper says so itself in its Scope note: it does not claim every component is universally optimal.

The Analogy: Practice Drawing and Taking Orders Separately

The prevailing paradigm (§1) shows an apprentice artist only pictures that come with an order form attached. The order form is the caption. Two problems follow. Order forms are expensive. And order forms can become lies — when a high-resolution photo is squeezed down to $256\times256$ for cheap early training, the fine details the caption mentions simply vanish from the image.

LLaDA-Image's move is simple: you don't need an order form to practice drawing. An image already contains the semantics needed to supervise its own generation, so a frozen vision-language model (VLM) can extract those semantics and hand them back as the conditioning signal (§4.2).

The Shape of the System: Three Parts, Two Paths

There are three components (§2). The understanding module is a diffusion-LLM-based VLM built on the LLaDA 2.0 Mini backbone. The connector is a two-stage bridge — a Residual Query Adapter (RQA) plus a shallow stack of Transformer blocks — that projects VLM representations into the DiT's conditioning space. The generator is a pure single-stream DiT: condition tokens and image tokens sit in one sequence and pass through the same blocks.

Two paths run through it. For text-to-image, the prompt travels RQA → frozen VLM → connector and arrives as the condition. Editing is the interesting case: the reference image never touches the VLM at all (§2.1). Its SigLIP-VQ features go through a DiT-specific branch and join the condition tokens, while a clean latent encoded by the FLUX.2 VAE is concatenated with the noised target latent (§2.3). The first path carries "what is in the picture"; the second carries pixel-level evidence for the regions that should not change.

Making an Image Its Own Teacher

Now the pre-training stage (§4.2). Feeding all the patch features in as conditioning would hand over essentially the whole answer, and the model would settle for copying its input. So a fraction $\rho_{\mathrm{img}}$ of the image tokens is masked out, turning the task from dense reconstruction into sparse-to-dense prediction. The objective is flow matching: draw noise $\mathbf{z}$ and a timestep $t\sim\mathcal{U}(0,1)$, form the interpolation $\mathbf{x}_t=(1-t)\mathbf{x}+t\mathbf{z}$, and minimize

Put in words: it is a regression that asks the model to name the direction from wherever it currently sits toward the clean image. Larger $t$ means noisier; $t$ near 0 means the picture has mostly formed. Only the DiT, RQA, and connector are updated here — the VLM and the vision encoder stay frozen.