Moonworks releases Lunara, a Diffusion Mixture Transformer for artistic image generation

Moonworks has released Lunara, a new image generation model built on a Diffusion Mixture Transformer architecture with fewer than 10B active parameters. The model is designed to model artistic intelligence, utilizing a novel CAT training algorithm that iteratively updates the training distribution through targeted sample acquisition, image refinement, and the selective inclusion of human-created artwork, inspired by active learning principles.

Evaluation and Performance

The model was evaluated using 1,000 shared prompts and 8,000 generated images, measuring aesthetic quality, emotional resonance, and content integrity against seven baselines: GPT-Image-1 Mini, Qwen-Image, AuraFlow, SD 3.5 Turbo, HiDream-I1 Fast, FLUX-Klein-4B, and Z-Image-Turbo.

• Automated Metrics: Under GPT-5.6 Sol evaluation, Lunara achieved the highest score for aesthetic quality at 8.473, surpassing GPT-Image-1 Mini (8.457) and Qwen-Image (8.366). However, GPT-Image-1 Mini led in emotional resonance and content integrity.
• Human Evaluation: In a blinded study with six evaluators assessing anonymized image pairs, Lunara achieved the highest mean scores across all three dimensions.

Resources

The release follows previous open-source dataset contributions by Moonworks. The paper is available on arXiv, and the evaluation dataset is hosted on Hugging Face.