stable-diffusion-image-generation

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

DPMSolverMultistepScheduler offers excellent quality in 15-25 steps and is a strong default. Use LCMScheduler for very fast 4-8 step generation, or EulerDiscreteScheduler as a reliable general-purpose choice.

Enable model CPU offloading with pipe.enable_model_cpu_offload(), add attention slicing and VAE slicing, and use float16 precision. For tighter budgets, enable sequential CPU offload or reduce image dimensions and batch size.

Yes, load LoRA weights with pipe.load_lora_weights() and adjust strength via fuse_lora(lora_scale=0.8). Multiple LoRAs can be combined using set_adapters with per-adapter weights.

Black images often come from the safety checker or VAE dtype mismatches—try disabling the safety checker or keeping the VAE in float32. Noisy output usually means too few inference steps or a misconfigured scheduler.

Use DALL-E 3 when you need API-based generation without managing GPU infrastructure. Stable Diffusion is preferable when you need local control, custom fine-tuning with LoRA or DreamBooth, or ControlNet spatial conditioning.