AuroraAI Research has published on Hugging Face AuroraIMG2-50M-Preview, a text-to-image model with a 50.4-million-parameter denoiser and a 12.6-million-parameter text encoder, released under the Apache 2.0 license. It is a preview version that generates fixed 256×256 pixel images and, according to the model card, was trained from scratch in ten hours on a single NVIDIA RTX PRO 6000.

Psst!Add AlexTech.ai asPreferred Source on Google

The model works in the latent space of the FLUX.2 VAE, the compressed representation where the image is built before being converted back to pixels. The final decoder is Black Forest Labs' pretrained FLUX.2 small decoder, included unchanged. The denoiser is a flow-matching transformer with eight blocks, while the text encoder is a six-layer transformer using an 8,192-token byte-level BPE tokenizer. Each pixel of the 16×16 latent grid becomes one token, with no patchify step.

The key figure concerns the scores. On GenEval 2 the model reaches 6.00 in Soft-TIFA geometric mean and 36.03 in arithmetic mean. The authors measured these with 50 sampling steps and guidance of 3.0, using fp32 weights with bfloat16 autocast. The card offers no comparison with other models on this benchmark, so the figure reads better as an internal measure than as a ranking.

The model card recommends 50 steps for quality and 20 for faster generation. No negative prompt is supported: the model uses the learned empty prompt, since arbitrary negative prompts were not part of training. The code requires trust_remote_code=True and the Diffusers library.

The release sits among other compact open models that run on modest hardware. The closest case is FLUX 3 Image Klein from Black Forest Labs, a 5.6-billion-parameter model that runs on an 8 GB card. The scale differs, but the aim is the same: moving image generation off the cloud. Lunara also sits in the under-ten-billion-parameter range, with evaluations focused on aesthetic quality.

According to agihunt.info, the end-to-end inference model totals 90 million parameters, including the text encoder and the pretrained VAE. The same sources indicate the pipeline is compatible with Diffusers and that the weights are distributed on Hugging Face.

For those who want to try it, the card shows direct loading through DiffusionPipeline with the custom_pipeline flag. The release does not yet include higher-resolution versions or independent evaluation on the most common benchmarks, and the Preview label signals the architecture may change in later versions.