Unlock full creative control with Stable Diffusion's open-source ecosystem, LoRAs, and ControlNet.
Stable Diffusion gives you capabilities no hosted platform can match: LoRA fine-tuning for specific styles or characters, ControlNet for precise composition control, inpainting for surgical edits, and unlimited generation at any resolution. The trade-off is a steeper learning curve and local hardware requirements (or cloud GPU costs). Mastering it means understanding not just prompts but model selection, sampling settings, and the LoRA ecosystem.
Each prompt is annotated with the reasoning behind its structure.
Positive: cinematic portrait of a young woman, soft natural window light, bokeh background, film grain, photorealistic, 8K, detailed skin texture, catchlights in eyes, shot on Sony A7III Negative: cartoon, anime, painting, illustration, deformed, ugly, blurry, low quality, watermark, text, duplicate, bad anatomy, worst quality, low resolution, extra limbs
Stable Diffusion's negative prompt is as important as the positive prompt. The negative prompt here blocks the model's tendency toward cartoon-style outputs when using base models. Specific camera gear (Sony A7III) activates photographic training data. Catchlights and skin texture are reliable quality anchors.
Prompt: [your character description] in a dynamic action pose, professional photography, dramatic lighting [add LoRA: <lora:your_character:0.8>] ControlNet: OpenPose — import a reference pose image. The AI will replicate the exact body position while generating your character in that pose.
ControlNet OpenPose extracts skeleton data from a reference image and applies it to the generation — you control the exact body position without describing it in text. Combining with a character LoRA lets you place a consistent character in any pose. The 0.8 weight prevents the LoRA from overpowering the prompt.
Positive: hyperrealistic architectural visualization, contemporary villa, glass and concrete, infinity pool overlooking Mediterranean sea, golden hour, volumetric light rays, photorealistic rendering, octane render quality, professional architectural photography Negative: cartoonish, low detail, people, cars, distracting elements, overexposed
SDXL (Stable Diffusion XL) handles architectural detail significantly better than SD 1.5. 'Octane render quality' anchors the style in 3D rendering training data. Removing people and cars in the negative prompt is common for clean architectural renders.
Original image generated. Mask: paint over only the sky. New prompt for masked area: dramatic stormy sky with cumulonimbus clouds, lightning, dark purple and orange, photorealistic. Denoising strength: 0.75. This replaces only the sky while keeping the foreground unchanged.
Denoising strength of 0.75 is the key parameter for inpainting — too high (0.9+) and it ignores your original image; too low (0.4-) and it can't make significant changes. Masking only the sky ensures the rest of the image stays pixel-perfect.
Node setup: Load Checkpoint (SDXL base) -> Load LoRA (character_v2, strength 0.75) -> Load LoRA (lighting_style, strength 0.5) -> KSampler (steps: 25, CFG: 7, sampler: dpmpp_2m, scheduler: karras) -> VAE Decode -> Save Image. Prompt: [character] in [scene], consistent lighting, same character from previous batch, seed: [fixed seed]
ComfyUI's node-based workflow lets you stack LoRAs with individual strength controls. Fixing the seed while changing the prompt produces consistent character variations. DPM++ 2M Karras at 25 steps is the most reliable sampler for quality/speed balance. Stacking two LoRAs (character + lighting) at reduced individual strengths prevents conflict.
Browse our full library of Stable prompts.
Browse Stable-diffusion Prompts