amd/Nitro-E: Lightweight text-to-image diffusion models
AMD researchers have introduced Nitro-E, a family of lightweight text-to-image diffusion models with 304 million parameters designed for efficient training and inference. The models utilize an Efficient Multimodal Diffusion Transformer (E-MMDiT) architecture featuring token reduction techniques, a compressive visual tokenizer, and Position Reinforcement to maintain spatial coherence. Training takes 1.5 days on a single node with eight AMD Instinct MI300X GPUs. Open article →