{"slug": "title-sdxl-performance-on-low-vram-gpu", "title": "Title: SDXL Performance on Low‑VRAM GPU", "summary": "Celestium Engine, an AI image generation tool, has added features to improve SDXL performance on low-VRAM GPUs, including dynamic FP16/FP32 switching in the VAE, autonomous VRAM cleanup, automatic CPU fallback, and the ability to run SDXL at 768x768 on 4GB GPUs with upscaling to 1536px using BuffaloCore, using around 3800 MB of VRAM.", "body_md": "Continuing the discussion from [SDXL Performance on Low VRAM](https://discuss.huggingface.co/t/sdxl-performance-on-low-vram/178826):\n\nHi, sharing the latest upgrades added to the Celestium Engine:\n\n**• Dynamic FP16 ↔ FP32 switching in the VAE**\n\nThe engine now automatically switches between FP16 and FP32 inside the VAE whenever higher detail stability is required.\n\nThis reduces artifacts and preserves image quality even on low‑VRAM hardware.\n\n**• Immediate and autonomous VRAM cleanup**\n\nEvery stage of the pipeline performs a full VRAM flush.\n\nZero fragmentation, zero accumulation, zero memory leaks.\n\nVRAM stays stable even after long multi‑generation sessions.\n\n**• Automatic Hardware Check**\n\nCelestium now detects when the GPU is unavailable or when VRAM is insufficient.\n\nIn those cases it automatically switches to **CPU fallback**, supported by system RAM, without interrupting the generation.\n\n**• Stable CPU fallback (consistent style & lighting)**\n\nWhen the GPU can’t continue, the engine preserves the same style, lighting, and visual coherence.\n\nNo aesthetic shift between GPU → CPU execution.\n\n**• SDXL starts at 768×768 even on 4GB GPUs**\n\nThe SDXL pipeline initializes at 768×768 with full stability and no crashes.\n\nWith BuffaloCore, upscaling reaches **1536px** even on 4GB GPUs, with real VRAM usage around **3800 MB**.\n\n**• Optimized BuffaloCore**\n\nDynamic tensor resizing prevents overflow and fragmentation.\n\nFully functional even on borderline hardware.", "url": "https://wpnews.pro/news/title-sdxl-performance-on-low-vram-gpu", "canonical_source": "https://discuss.huggingface.co/t/title-sdxl-performance-on-low-vram-gpu/179054#post_1", "published_at": "2026-08-21 07:36:11+00:00", "updated_at": "2026-08-21 12:12:53.965736+00:00", "lang": "en", "topics": ["generative-ai", "ai-tools", "ai-infrastructure"], "entities": ["Celestium Engine", "SDXL", "BuffaloCore"], "alternates": {"html": "https://wpnews.pro/news/title-sdxl-performance-on-low-vram-gpu", "markdown": "https://wpnews.pro/news/title-sdxl-performance-on-low-vram-gpu.md", "text": "https://wpnews.pro/news/title-sdxl-performance-on-low-vram-gpu.txt", "jsonld": "https://wpnews.pro/news/title-sdxl-performance-on-low-vram-gpu.jsonld"}}