The base model is Wan 2.2 T2V 14B low noise. The prompts are generated with JoyCaption beta one.There are two versions of this LoRA. One is trained with shift = 1 to create the details (for example, how to draw plants and water). The other is trained with shift = 12 to create the overall composition (for example, the horizon line should be at 1/2 or 1/3 from the top so there is some space for the dialogue box at the bottom, and there can be an airbrush overlay layer for the sunlight).(Now I realized shift = 12 may be too high. Some say Wan actually used shift = 5 when training.)I recommend to set shift = 3 when generating. It's ok to use only the shift1 LoRA. For better result, you may use the shift12 LoRA for the first few steps, then use the shift1 LoRA for the remaining steps (see the workflow in the example images).You can use a Canny ControlNet or any recent image editing models to generate variations with the same objects and different lightings.

Files with SHA256: e0bc478901de9a0ba4d6109beb7ac68ef84f36450b27bbd00bdc15d6673a8494