NightCafe exposes several advanced settings for fine-tuning image generation. You don’t need these for great results, but they give extra control when you want it.
Enabling CLIP Guidance can improve results for complex prompts or larger resolutions, and can make images look more realistic. It changes the sampling method behind the scenes.
Available for Stable Diffusion 1.5 base models only, CLIP Skip skips layers of the text-interpretation model, which can give broader or more specific results. It’s an advanced setting — only use it if you’re familiar with the model.
The SDXL pipeline has a base model plus a refiner that adds detail and vividness. Refiner weight is the percentage of steps given to the refiner (default 50%). Some checkpoints — e.g. DreamShaper and Juggernaut — look best with a low or off setting, so experiment based on the style you’re after.
Raw mode produces a less synthetic, more natural aesthetic — closer to candid photography. It increases diversity in human subjects and boosts realism for nature photography. Turn it on when you want authenticity over a polished, “AI” look.
Grounded search improves accuracy by searching the web for real-world reference images related to your prompt, helping the model render specific subjects, objects or scenes more faithfully. It adds a small credit surcharge (1.5 credits at the time of writing) to the base cost of your generation.