More specific than feels necessary. Naming four things — arms, torso direction, weight-bearing leg, head — reliably produces a real change. Asking for a better pose or a more natural pose usually returns something almost identical to the original, because nothing in that request tells the model what to move.

Generally yes, and it is the part to check first on every run. The face is copied through the edit rather than regenerated, so it holds up well from a clear front-facing source. Where identity drifts is when the new stance turns the head far away from the camera and the model has to invent a profile it has never seen.

It can, and hands are still the weak point — this is worth knowing before you plan a shot around them. Fingers wrapped around an arm or interlocked in front of the body come out wrong often enough that it is worth checking closely and rerunning. Simpler positions, such as arms folded or hands in pockets, land far more reliably.

A plain or softly blurred background survives cleanly. A detailed one — shelves, tiles, a patterned wall, text on a sign — is where you see the repair, because the area the body used to cover has to be reconstructed from what surrounds it. Simple backdrops are the reason studio-style shots restage best.

Not by uploading that picture as a second reference here — this flow takes one image and a description. What works instead is describing the stance in that reference in words: the angle of the shoulders, where the hands sit, which way the hips face. The description is the interface.

It should not, and it is not meant to. The instruction is about position, not proportions, so build, height and body shape are carried through. If they shift, the prompt usually asked for a shape word alongside the stance — remove it and rerun.

The estimate appears before you generate and moves with the format and resolution you pick, so the cost is visible first. Failed runs are never charged, and new accounts start with free credits — enough to try a second stance before spending anything.