Stable Diffusion

SD 1.5 is the original open-weights Stable Diffusion release from Runway and Stability AI, and it is the model that started the community. It is a latent diffusion model built on the CLIP text encoder with a native resolution of 512×512, which is why it is tiny to run and lightning-fast — it fits comfortably on GPUs with well under 10GB of VRAM, or none at all when you run it here. Rather than reading long natural-language descriptions, it works from short, comma-separated tag prompts, the same style that shaped years of community workflows.

Almost nobody runs the raw base checkpoint today — the ecosystem lives in its fine-tunes. The default hosted here is DreamShaper, a long-running community model whose author set out to build "a better Stable Diffusion," a versatile "swiss-knife" checkpoint aimed first at art and illustration. Around SD 1.5 sits the deepest support stack of any open model: textual inversions, LoRAs, ControlNet, negative embeddings like the DreamShaper author’s "Bad Dream," and img2img / highres-fix upscaling pipelines that push its 512-native output to higher resolutions.

Choose SD 1.5 when speed, low hardware cost, and sheer breadth of styles matter more than raw prompt fidelity — it renders in a second or two and has the largest LoRA and fine-tune library anywhere. For higher native resolution and cleaner anatomy, the SDXL-based ecosystems (SDXL, Pony, Illustrious) are the natural step up, and Flux leads on prompt adherence and in-image text. But for fast iteration, stylized art, and reusing a decade of community resources, SD 1.5 remains the lightest and most flexible starting point.