SupraLabs debuts Supra2-IMG, a 100M-parameter image generator
SupraLabs just open-sourced Supra2-IMG, a tiny diffusion transformer with only 104M parameters and a footprint around 417MB. It was trained on 5.6M synthetic images in just 9 hours on a single H100 GPU. While resolution is capped at 256x256, it proves practical text-to-image models can comfortably squeeze onto phones and edge devices.
Supra2-IMG is an ultra-compact text-to-image model sporting just 104M parameters, packed into a file size of around 417MB. It is built to run directly on smartphones, tiny devices, and edge hardware without relying on the cloud.
While mainstream image generators usually tip the scales at multiple billions of parameters, SupraLabs managed to shrink the footprint by dozens of times, testing the limits of what a tiny architecture can achieve from scratch.
Built from scratch in 9 hours
Supra2-IMG is a tiny Diffusion Transformer (DiT) trained entirely on synthetic data. Instead of downsizing an existing giant checkpoint, the team started training from the ground up.
It leverages a frozen Flan-T5-Base encoder with a 128-token context window to interpret text prompts, paired with Stability AI's classic SD-VAE-FT-MSE for decoding latent patches.
The training pipeline processed 5.6M images selected from the FLUX-Reason-6M dataset over 10 epochs. Prompts were strictly filtered by composition, entity, text, style, and imaginative captions to keep data quality high.
Remarkably, the entire training run happened on a single Nvidia H100 SXM 80GB pod on Runpod. Including data preparation, training took only 9 hours and finished with an FID score of 27.9 and a CLIP score of 29.4.
Running the model and its trade-offs
Getting it up and running takes just a couple of lines using Hugging Face Transformers, or you can grab the official inference.py script directly.
Recommended inference settings use Euler flow at 50 steps with a CFG value of 3.0. Loading the entire model into memory takes less than a second.
You should keep expectations grounded, though. Due to its miniature size, native output is fixed at 256x256 resolution. Complex prompts and fine-grained textures can look muddy compared to modern behemoths.
The weights are fully available on Hugging Face alongside ready-to-run Google Colab and Kaggle notebooks.
Why it matters
We think Supra2-IMG is a fascinating proof of concept for local and edge computing. If you are an app developer or IoT hacker looking for an offline, featherweight visual generator, downloading a 417MB model is completely painless.
It highlights how far clean synthetic data can carry a budget model. It won't replace your desktop generative workflow, but for embedded prototyping and on-device experiments, it is well worth playing with.
Sources (1)
- huggingface.co(opens in a new tab)Originalhuggingface.co