Tsubaki.3 test: per-face edits beat batch sheet, under 5 min

The test starts from a practical problem: batch generation gives you six expressions with few operations, but a sheet is a single image, so any correction is applied to the whole thing. The author kept one ordinary-expression reference image fixed throughout and avoided chaining outputs, noting that using a generated expression as the next reference can let faces and hairstyles drift. In the batch route, the sad face came out weak, and a prompt naming the bottom-left position and spelling out brow, eye and mouth changes did fix it — yet other panels shifted anyway.

The comparison table the author built shows where the time went. The batch route involved generation, partial correction and a full recheck, with the third attempt finally acceptable. The individual route produced the normal face plus five expressions separately, and the high-speed mode kept the whole set under 5 minutes. Because a finished smile is a separate file, reworking the sad face cannot disturb it.

The author is careful about the headline number: the individual edits used high-speed mode, so the gap should not be read as a pure difference between the two methods, and timing also varies with settings and congestion. The takeaway offered is a split workflow — use batch generation to review expression direction for reference material, then finish only the assets actually needed one at a time.