Native low-bit (NLT) diffusion models — the weights live in the quantized space from step 0 instead of being compressed after fp16 pre-training. Home of the AquariusImage series: Aquarius Terimage (ternary text-to-image, released) and Aquarius Binimage (planned).
research pytorch image-generation gradio text-to-image from-scratch model-compression efficient-inference diffusion-model deep-learning-research ternary-quantization quantization-aware-training low-bit-quantization int4 stable-diffusion generative-ai native-low-bit-training
-
Updated
Oct 4, 2026 - Python