The Celestium Engine has introduced several upgrades to improve SDXL performance on low-VRAM hardware. These changes focus on memory management, stability, and fallback mechanisms to ensure consistent generation even on constrained systems.

  • Dynamic FP16 ↔ FP32 switching in the VAE reduces artifacts and preserves image quality by automatically adjusting precision based on detail stability requirements.
  • Immediate and autonomous VRAM cleanup performs a full flush at every pipeline stage, eliminating fragmentation, accumulation, and memory leaks.
  • Automatic Hardware Check detects insufficient GPU or VRAM conditions and switches to CPU fallback supported by system RAM without interrupting generation.
  • Stable CPU fallback maintains consistent style, lighting, and visual coherence, preventing aesthetic shifts between GPU and CPU execution.
  • SDXL initialization starts at 768×768 with full stability on 4GB GPUs, while BuffaloCore upscaling reaches 1536px with real VRAM usage around 3800 MB.
  • Optimized BuffaloCore uses dynamic tensor resizing to prevent overflow and fragmentation, ensuring functionality on borderline hardware.

These updates allow users to run SDXL pipelines on lower-end GPUs or fall back to CPU execution without losing visual quality or encountering crashes.