Mia Lindgren

My split is between the prompt and the provenance. The prompt is the part everyone tunes. The provenance is the part that determines whether the file can be used. A useful prompt states subject, action, setting, composition, lighting and style, and the order does not matter much. What matters is that vague terms get interpreted differently each time, so the same prompt rarely produces the same image. Where reproducibility matters, a fixed seed plus a recorded model version and settings is the closest thing to a guarantee, and it is weaker than it sounds because model updates change the mapping. Aspect ratio is set at generation rather than cropped afterwards, and asking for a composition that depends on the frame will fight a later crop. That is a composition problem rather than a technical one, and it saves work to decide before generating. Metadata is where I get specific. Generated images carry fields describing the tool and settings used, and photographs carry much more: camera, lens, timestamp, and frequently location. The provenance standard for generated content exists because a file claiming to be a photograph when it is not has caused real harm in news contexts. Upscaling enlarges without inventing detail, which is the correct expectation to set. Extending a frame generates new content at the edges, which is a different operation with different failure modes, mainly at the seams. I keep licensing notes on every page, because the terms differ by tool and change without much notice. One last practical note. Ask for a batch and pick from it, rather than generating one at a time until something works. Generation is the slow part, and the same prompt usually gives you three usable variants and one you keep.