Why outputs vary from run to run
The model fabricates a plausible body for every run — there's no single "correct" answer it's converging on. Two runs on the same photo can differ in anatomy detail, skin shading and overall plausibility.
Small input changes move the output too: a tighter crop or a different photo from the same set can produce a noticeably different result. When a run misses, retrying is normal, not a bug.