SDXL
SDXL (Stable Diffusion XL) is a large-scale text-to-image generation model released by Stability AI and an upgraded version of the Stable Diffusion series. It employs a dual text-encoder architecture with a substantially larger parameter count, enabling the generation of higher-resolution images with richer detail. The model also demonstrates stronger semantic understanding of complex prompts and is widely used in AI art creation and design assistance applications.
Timeline (last 90 days)
By default, after generating one image, the SDXL checkpoint, CLIP text encoder, and VAE remain resident in VRAM, occupying about 8GB; after calling /free, VRAM falls back to a baseline of about 2.2GB.
In ComfyUI, running SDXL alone to generate one image takes about 15 to 30 seconds.
An Ollama 14b model combined with SDXL 1024×1024 generation requires approximately 14.5 GB of VRAM, which is tight but barely feasible on a 16 GB card.
SDXL generating a 1024×1024 image occupies approximately 6–8 GB of VRAM.
SD XL introduced a dual-encoder architecture (CLIP ViT-L + OpenCLIP ViT-bigG), improving semantic capture for long prompts.
The SDXL base model alone occupies approximately 6–7 GB of VRAM, and peak VRAM can easily exceed 20 GB when simultaneously loading the Refiner, multiple ControlNet weights (approximately 1.4 GB each), and a high-precision VAE.
According to the publisher's description, Nanosaur 2 outperforms the SDXL model in many comparisons, despite having far fewer parameters than SDXL.
The score gaps between subsequent models such as Flux and SDXL on public benchmarks have already become quite limited.
Running mainstream image generation models (such as SDXL) requires at least 8GB of VRAM
CFG=4–4.5 is the recommended starting range for many Flux and SD3 series models, while SDXL-type models commonly use 5–8.
19 more timeline events
All Facts (20)
SDXL采用了双CLIP编码器（OpenCLIP ViT-bigG和OpenAI CLIP ViT-L）的设计
90%Unverifiedsd-scripts支持LoRA、LyCORIS（LoHa、LoKr等）、Textual Inversion、DreamBooth等微调方法，并支持SD 1.5、SDXL和SD3
90%UnverifiedSD1.5使用768维条件向量，SDXL使用1024维条件向量
85%UnverifiedSD 1.5系列模型约需4-6G显存，SDXL系列通常需要8-12G显存
85%UnverifiedBetween 2023 and 2024, the image generation field experienced an architectural leap from Stable Diffusion 1.5 to SDXL and then to SD3.
60%UnverifiedSDXL（Stable Diffusion XL）是Stability AI推出的高分辨率图像生成模型，其完整管线包含Base模型和Refiner模型两个阶段
60%UnverifiedIn ComfyUI, running SDXL alone to generate one image takes about 15 to 30 seconds.
50%UnverifiedBy default, after generating one image, the SDXL checkpoint, CLIP text encoder, and VAE remain resident in VRAM, occupying about 8GB; after calling /free, VRAM falls back to a baseline of about 2.2GB.
50%UnverifiedAn Ollama 14b model combined with SDXL 1024×1024 generation requires approximately 14.5 GB of VRAM, which is tight but barely feasible on a 16 GB card.
50%UnverifiedSDXL generating a 1024×1024 image occupies approximately 6–8 GB of VRAM.
50%UnverifiedSD XL introduced a dual-encoder architecture (CLIP ViT-L + OpenCLIP ViT-bigG), improving semantic capture for long prompts.
50%UnverifiedThe SDXL base model alone occupies approximately 6–7 GB of VRAM, and peak VRAM can easily exceed 20 GB when simultaneously loading the Refiner, multiple ControlNet weights (approximately 1.4 GB each), and a high-precision VAE.
50%UnverifiedAccording to the publisher's description, Nanosaur 2 outperforms the SDXL model in many comparisons, despite having far fewer parameters than SDXL.
50%UnverifiedThe score gaps between subsequent models such as Flux and SDXL on public benchmarks have already become quite limited.
50%UnverifiedRunning mainstream image generation models (such as SDXL) requires at least 8GB of VRAM
50%UnverifiedCFG=4–4.5 is the recommended starting range for many Flux and SD3 series models, while SDXL-type models commonly use 5–8.
50%UnverifiedSDXL has approximately 6.6B parameters.
50%UnverifiedPrompt Studio supports customized prompt syntax optimization for different models including Generic, FLUX, Krea 2, and SDXL.
50%UnverifiedFor those seeking a balance between speed and quality without pursuing the absolute best, SDXL and its derivative models are the most fitting choice.
50%UnverifiedSDXL typically achieves good results in 25–35 steps; blindly increasing the step count yields diminishing returns and slows down generation.
50%