Moonworks' Lunara: Sub-10B Diffusion Mixture Transformer Tops Aesthetic and Human Blind Evaluations
paper-crow · reddit · 2026-10-09
Moonworks released the Lunara paper and open eval dataset:
- Architecture: a Diffusion Mixture Transformer with <10B active parameters, targeting "artistic intelligence" in image generation.
- Training: the CAT algorithm, inspired by active learning, iteratively updates the training distribution via targeted sample acquisition, image refinement, and selective inclusion of human-created artwork; semantic variations build controlled neighborhoods of related examples.
- Evaluation: 1,000 shared prompts, 8,000 images, seven baselines (GPT-Image-1 Mini, Qwen-Image, AuraFlow, SD 3.5 Turbo, HiDream-I1 Fast, FLUX-Klein-4B, Z-Image-Turbo). Under GPT-5.6 Sol judging, Lunara leads aesthetic quality at 8.473 vs 8.457 (GPT-Image-1 Mini) and 8.366 (Qwen-Image); GPT-Image-1 Mini leads emotional resonance and content integrity.
- Blind human eval: six evaluators scored anonymized pairs; Lunara topped all three dimensions.
Paper: arxiv.org/abs/2609.22272; dataset on Hugging Face.
More from Research
- RWTH Aachen's ARROW unifies 3D reconstruction and point tracking from any RGB inputs, sets new SOTA — kwangmoo_yi · 2026-10-09
- HAIPS 2026 workshop lands at COLM tomorrow with top AI privacy researchers — tianshi_li · 2026-10-09
- Tighter Bayes error bounds via generalized Bhattacharyya and Chernoff means — FrnkNlsn · 2026-10-09
- PAMI: part-anchored motion lifts contact recall 14.5% in text-to-HOI generation — _akhaliq · 2026-10-09
- Study finds LLM novelty judges are unstable: scores swing wildly with evaluation design choices — _akhaliq · 2026-10-09
- Google's AMIE lands in The Lancet: 90% match with doctors' final diagnoses in real clinics — sundarpichai · 2026-10-09