October 9, 2026
Lunara beats rivals in human ratings: first place across 3 criteria
In Moonworks' Sep 11 study, people rated Lunara above its competitors for beauty, emotional impact and image coherence.
GPT-5.6 Sol picked a different leader on two criteria. It ranked GPT-Image-1 Mini above Lunara for emotional impact and image coherence: how consistently the details fit together.
Moonworks compared Lunara with 7 other AI models on 1,000 prompts. GPT-5.6 Sol wrote all the prompts itself and evaluated 8,000 images with the model names hidden.
Nearly tied on beauty. In the authors' test, Lunara beat GPT-Image-1 Mini by 0.016 points. Below are GPT-5.6 Sol's ratings from the Sep 11 study.
| Criterion | Lunara | GPT-Image-1 Mini | | --- | --- | --- | | Beauty | 8.473 | 8.457 | | Emotional impact | 8.119 | 8.363 | | Image coherence | 7.944 | 8.609 |
In a blind comparison, 6 people gave Lunara the highest average ratings on all three criteria. They rated Lunara 1,000 times and each competitor 142–143 times.
Moonworks trains Lunara by repeatedly selecting and refining examples. The CAT algorithm adds suitable images, including human-created work, and generates variations with a shared theme but different compositions.
Where to try Lunara. As of Oct 8, Moonworks offers the model in the Echoes beta. The authors claim it generates a 1024×1024 image in under 10 seconds on an A100 with 40 GB of memory.
Sources: [Moonworks' Sep 11 study](https://arxiv.org/abs/2609.22272), [Echoes beta](https://echoes.moonworks.ai/), [Lunara evaluation dataset](https://huggingface.co/datasets/moonworks/lunara-art-eval).
Moonworks published 8,000 test images along with prompts, model names and scores across three criteria.