creating-assets

Paste the same style paragraph verbatim at the head of every prompt, never paraphrased. For things that must look identical, like one hand in six poses, generate them together in a single image and cut them apart afterwards rather than generating separately.

Let the assistant write the prompts because it knows what the code needs, such as separate files for parts that move and matching pairs generated together. Spend as much of the prompt on what not to do, like no gradients or shadows, as on what to make.

Yes, supplied images are used as given without recolouring or filtering, and are placed in the same figure area with objectFit contain. Check they are at least roughly 800px wide for a 1920x1080 frame, and store them under public/assets/user-uploads.

Inconsistency almost always comes from a missing or paraphrased style preamble, or from generating matching subjects separately. Also never pass photographs or others' artwork as style references, since generators copy what they are shown.

Draw in code when the subject is an idea rather than a thing, or when accuracy matters more than richness, such as diagrams with labelled parts. Code-drawn shapes are exact, re-render at any size, cost nothing, and are edited by changing a number.

Generate at least twice the size the picture appears on screen, with a fully transparent background and no shadow baked in. Scaling down looks fine while scaling up does not, and shadows belong in code where they can move.