Celestium pipeline enables SDXL on 4GB-8GB GPUs via adaptive FP16 and CPU fallback

A diagnostic using Celestium’s hardware-agnostic inference pipeline demonstrates that Stable Diffusion XL can run stably on low-VRAM GPUs (4GB, 6GB, and 8GB) by leveraging adaptive FP16 precision and autonomous memory cleanup. The approach prevents fragmentation spikes during multi-stage pipelines and utilizes an emergency CPU mode to maintain generation when GPU VRAM is exhausted.

• SDXL Base + Refiner runs on 8GB GPUs, while SDXL Base only is recommended for 4GB cards.
• Adaptive FP16 precision reduces VRAM load, and BuffaloCore dynamically resizes tensors to avoid overflow.
• RealESRGANx2 adds minimal overhead even on 4GB GPUs, and the system requires zero cloud or CUDA dependencies.

This method allows SDXL to run offline and locally without crashes on consumer hardware that normally struggles with the model's memory requirements.