# Title: SDXL Performance on Low‑VRAM GPU

> Source: <https://discuss.huggingface.co/t/title-sdxl-performance-on-low-vram-gpu/179054#post_1>
> Published: 2026-08-21 07:36:11+00:00

Continuing the discussion from [SDXL Performance on Low VRAM](https://discuss.huggingface.co/t/sdxl-performance-on-low-vram/178826):

Hi, sharing the latest upgrades added to the Celestium Engine:

**• Dynamic FP16 ↔ FP32 switching in the VAE**

The engine now automatically switches between FP16 and FP32 inside the VAE whenever higher detail stability is required.

This reduces artifacts and preserves image quality even on low‑VRAM hardware.

**• Immediate and autonomous VRAM cleanup**

Every stage of the pipeline performs a full VRAM flush.

Zero fragmentation, zero accumulation, zero memory leaks.

VRAM stays stable even after long multi‑generation sessions.

**• Automatic Hardware Check**

Celestium now detects when the GPU is unavailable or when VRAM is insufficient.

In those cases it automatically switches to **CPU fallback**, supported by system RAM, without interrupting the generation.

**• Stable CPU fallback (consistent style & lighting)**

When the GPU can’t continue, the engine preserves the same style, lighting, and visual coherence.

No aesthetic shift between GPU → CPU execution.

**• SDXL starts at 768×768 even on 4GB GPUs**

The SDXL pipeline initializes at 768×768 with full stability and no crashes.

With BuffaloCore, upscaling reaches **1536px** even on 4GB GPUs, with real VRAM usage around **3800 MB**.

**• Optimized BuffaloCore**

Dynamic tensor resizing prevents overflow and fragmentation.

Fully functional even on borderline hardware.