Inclusion AI has unveiled LLaDA-Image, a text-to-image model that incorporates a comprehensive diffusion architecture. At its heart lies a 6-billion-parameter DiT (Diffusion Transformer), meticulously trained from the ground up. A unified set of weights can seamlessly facilitate both text-to-image generation and instruction-based image editing functionalities. Moreover, this model can be distilled into LLaDA-Image-Turbo, a variant that necessitates only 2-4 sampling steps for operation. LLaDA-Image employs a distinctive training strategy: "painting first, then comprehending language." Throughout the pre-training and mid-training stages, a substantial majority—over 90%—of the cumulative 220 million training samples were supervised using pure images to establish a robust visual foundation. Subsequently, during the SFT (Supervised Fine-Tuning) phase, validated image-text pairs were utilized to achieve alignment with language. The model secures the top position among open-source models on both the Chinese and English leaderboards of Qwen-Image-Bench, demonstrating exceptional performance in tasks such as text generation and instruction-based image editing. Furthermore, it ensures training stability through the use of parameter-free RMSNorm and leverages TwinFlow technology for swift distillation. Presently, the model weights, training code, and comprehensive recipe for LLaDA-Image have been fully open-sourced, and the model has been adapted to seamlessly integrate with the SGLang inference framework.