SupraLabs Releases Compact Supra2-IMG Text-to-Image Model

The model SupraLabs/Supra2-IMG has been published on HuggingFace — a diffusion transformer for generating images from text descriptions. The author calls it tiny: about 100 million parameters, trained from scratch on synthetic data, output resolution 256x256.

What it means

The largest weights file — model_final_ema.pt — takes up 0.4 GB, which is noticeably less than typical text-to-image models. The apache-2.0 license permits free use and modification. The model runs via transformers, and the weights are published in pt format. The architecture is a diffusion transformer with a frozen Flan-T5-Base encoder and a VAE from Stability AI. The developer does not specify exact VRAM requirements, but with 104 million parameters and a 0.4 GB weights file, the model should fit into VRAM volumes that are modest by today's standards. The model was trained on 5.6 million images from the FLUX-Reason-6M dataset over 10 epochs on a single H100. For those looking for a lightweight text-to-image tool for experiments on their own hardware, this is one of the most compact options among diffusion transformers. You can follow the release of similar models in the news feed about AI tools.