Technical Specifications and Training Efficiency
SupraLabs has announced the release of Supra2-IMG, a text-to-image model characterized by its compact architecture. According to a technical announcement shared on Reddit, the model contains 100 million parameters and utilizes a Diffusion Transformer (DiT) architecture.
The developers report that the model was trained from scratch in under 10 hours using a single NVIDIA H100 GPU on the Runpod platform. The model is designed to generate images at a resolution of 256x256 pixels. While the developers claim the model achieves “state-of-the-art” quality, some community feedback on Reddit has noted the presence of diffusion artifacts in the generated samples.
Accessibility and Local Deployment
The model has been made available via Hugging Face, allowing for local deployment on consumer-grade hardware. Performance benchmarks provided by the developers suggest significant differences in processing speeds depending on the hardware used: approximately 20 seconds per image on a CPU, and approximately 2 seconds per image on a GPU.
The release includes an inference script to facilitate local use. Users can generate images by specifying prompts, seeds, classifier-free guidance (CFG) scales, and step counts through a command-line interface.
Community Reception
Initial reactions from the developer community have been mixed. Some users highlighted the model’s potential for mobile applications due to its small size, while others expressed skepticism regarding the “state-of-the-art” claims made by the developers. Despite these critiques, the release has garnered significant engagement on social media platforms, with users expressing interest in the model’s efficiency and requesting further information regarding upscaling techniques to reach higher resolutions.