Nucleus-Image-FP8-Native

This is a native FP8 (float8_e4m3fn) quantization of the 17B Nucleus-Image model.

⚠️ VRAM Warning

This model is EXTREMELY HEAVY. Even in FP8, the weights alone take up ~13-14GB of VRAM.

• 16GB VRAM (RTX 4060 Ti / 4070 Ti Super / 4080): Recommended to use with sequential_cpu_offload for stability. Pure GPU inference might OOM at 1024x1024.
• 24GB VRAM (RTX 3090 / 4090): Best experience. Can run Pure GPU without offloading.
• 12GB VRAM or less: NOT RECOMMENDED unless using heavy CPU offloading (will be slow).