Using ControlNet with Stable Diffusion, I used your input image first transform into lineart version to maintain the composition. Then provide specific prompts (like "dog") to guide the model, producing the desired image output. It is really powerful and flexible. Whole process takes me 1 minute.