I use the SDXL_VAE and SDXL Offset-lora to condition my latent with two sdxl refiner prompts and ksamplers with lighting and illumination descriptions for 4 steps and blend them into one latent to feed into the k-sampler of my Base Image prompt that starts at step 5.  The resulting image from what I call the conditioning workflow adds guides for the text to image to follow influencing the final rendered images, sometimes dramatically and sometimes a little more subtle.