stable-diffusion

Load a DiffusionPipeline from a pretrained model like stable-diffusion-v1-5 with torch.float16, move it to CUDA, and call it with a prompt, num_inference_steps, and guidance_scale. The returned images list contains PIL images you can save directly.

SD 1.5 generates 512x512 images with lower VRAM needs, SDXL produces higher quality 1024x1024 outputs, and Flux uses newer transformer-based architectures. Diffusers auto-detects the model type when loading with DiffusionPipeline.

Enable model CPU offloading with pipe.enable_model_cpu_offload(), add attention slicing and VAE slicing, or load the model in float16 precision. Reducing image dimensions and batch size also lowers memory usage.

Yes, Diffusers provides AutoPipelineForImage2Image for transforming existing images with a strength parameter, and AutoPipelineForInpainting for filling masked regions defined by a white mask image with text-guided content.

Call pipe.load_lora_weights() with the LoRA path and weight name, then optionally fuse with pipe.fuse_lora(lora_scale=0.8) to adjust strength. Multiple LoRAs can be combined using set_adapters with per-adapter weights.

Default schedulers need 50 steps; switch to DPMSolverMultistepScheduler for 20-step generation or LCMScheduler with the LCM LoRA for 4-step results. Enabling xFormers attention and torch.compile also speeds up inference.