SupraLabs Releases Supra2-IMG, a Lightweight 100M Parameter Text-to-Image Model
SupraLabs has released Supra2-IMG, an open-source 100-million-parameter Diffusion Transformer (DiT) model capable of generating 256x256 resolution images. The model was trained entirely from scratch in under 10 hours on a single NVIDIA H100 GPU. The release demonstrates remarkable compute efficiency for text-to-image generation, showing that compact DiT architectures can produce coherent image samples at minimal training costs. It allows developers to run fast local image generation directly on consumer hardware, including CPUs without dedicated GPUs. Supra2-IMG generates images locally in roughly 2 seconds on a GPU and around 20 seconds on a CPU using 50 sampling steps. However, its main limitation for real-world applications is its restricted output resolution of 256x256 pixels.
## BACKGROUND
Diffusion Transformers (DiT) replace the conventional U-Net backbone in image diffusion models with scalable transformer architectures. Generative diffusion models often rely on Classifier-Free Guidance (CFG) during inference to adjust how closely the synthesized visual output matches the input text prompt.