LLaDA-Image: A Fully Open 6B Image Generator

inclusionAI has released LLaDA-Image, a 6-billion-parameter diffusion transformer for text-to-image generation and image editing, along with a rare full accounting of how it was built. According to the accompanying paper, the project's central claim is not just the model itself but a completely open training recipe — the kind of detail most labs keep behind closed doors.

The architecture leans on a frozen vision-language model rather than training a bespoke text encoder from scratch. That design keeps the language-understanding component fixed while the diffusion transformer learns to turn prompts into pixels, a strategy meant to make strong results reproducible without an enormous compute budget.

Why it matters

Open weights are common; open recipes are not. By documenting the data pipeline, training procedure, and design choices end to end, LLaDA-Image gives researchers and smaller teams a template they can actually follow and modify.

• 6B diffusion transformer covering both text-to-image and image editing
• Frozen VLM for prompt understanding, reducing training complexity
• Distilled few-step variant for faster inference

The distilled version is the practical hook here: few-step sampling cuts the generation latency that makes diffusion models expensive to serve. Combined with a mid-sized parameter count and a permissive-enough footprint, LLaDA-Image positions itself as a workable base for teams that want to study or extend open image generation rather than treat it as a black box.