# Qualitative inspection / 定性检查

Agent visual inspection of the seed-42 pairs; not a blinded human preference study,
OCR benchmark, or quantitative semantic score. All seed-123 outputs remain available
in the gallery for independent review.

| Case | Observed BF16 / INT8 behavior |
|---|---|
| Portrait | Both preserve the requested older woman, blue sweater and window composition. Face shape, wrinkles, framing and fabric details shift slightly. Neither is pixel-identical. |
| English lettering | Both visibly render “CREATE WITH LIGHT” and “September 2026”. Layout and desk objects differ. |
| Chinese lettering | Both visibly render the requested main title and subtitle. Both also add unwanted, partly nonsensical footer text despite the instruction to add no other text. |
| Counting / arrangement | Both show three cups in red/blue/yellow order with one green apple in front of the blue cup. Object scales and spacing differ. |
| Texture | Both depict a plausible blue/orange kingfisher. Head angle, branch orientation and feather texture differ noticeably. Low pixel similarity here is not by itself evidence of poor semantic quality. |
| RGBA | Both produce a green/orange dragon sticker and a real varying alpha channel. Sampled background pixels have alpha 0–1/255; the subject is opaque. Invisible RGB background colors should not be mistaken for an opaque background. Both have a white sticker outline. |
| Editing | Both recolor the sweater toward a red/burgundy tone, but produce conspicuous oversharpening and contrast/texture changes across the portrait and background. Faithful preservation of lighting and skin texture is not achieved in this test. |

**Editing limitation:** The artifact is reproducible in both precisions under this
specific pinned runtime and test setup. This comparison does not identify its root
cause and does not attribute it specifically to INT8. Do not treat successful pipeline
execution as validation of editing fidelity.

中文要点：文生图整体语义在所检查样例中保留，但画面细节会变化。中英文主标题可读；
中文海报的多余页脚在两种精度中均存在。编辑任务在两组都出现明显过度锐化及对比度变化，
不能宣称“编辑无损”或“质量完全一致”。这些观察只针对所列样例，不代表通用榜单结论。