stable-diffusion-image-generation

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

DPMSolverMultistepScheduler gives high quality in 15-25 steps and is a strong default. LCMScheduler enables 4-8 step generation for speed, while EulerDiscreteScheduler is a reliable general-purpose choice at 20-50 steps.

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

Yes, call pipe.load_lora_weights() with the LoRA path and weight name, then optionally fuse with pipe.fuse_lora(lora_scale=0.8). Multiple LoRAs can be combined using named adapters and pipe.set_adapters() with per-adapter weights.

Black images often come from the safety checker or VAE dtype mismatches; try disabling the safety checker and keeping the VAE in float32. Noisy outputs usually mean too few inference steps or a misconfigured scheduler.

SD 1.5 generates 512x512 images with lower memory needs, SDXL produces higher-quality 1024x1024 images with a larger model, and Flux uses a transformer-based architecture via FluxPipeline. Each has a dedicated pipeline class in Diffusers.