5 次定向修订：失败首图 → 实际编辑词 → 最终候选

以下编辑指令从当次图像工具实际调用中恢复，英文正文未改写。私人文件路径未发布，改为本库可点击的同一张 v1 首图。03 当时使用最近一张图引用，其余四次使用本地图片路径引用；五次都只引用一张。

5 张失败首图现已全部收录在 before/，与 12 张最终候选分别计数。 你可以打开原始输入图片并重用下方实际编辑词；模型构建标识和固定 seed 未归档，所以不保证重新编辑获得相同像素。这些都是本实验自己的生成图，没有第三方图库或额外生图。

首图保留为 rejected_initial_edit_input，最终结果保留为 candidate / internal；用户接受仍待确认。每例对照图都可点击查看完整原图。

03 prop-asset

已观察缺陷：six speaker-hole rows instead of three。

失败首图 · 编辑输入	最终候选 · 编辑输出

打开 v1 编辑输入 → 使用下方实际编辑词 → 查看 v2 最终候选

输入图片 SHA-256：62bdf30886f36e0a10acaf6958c928aaa500f01cf8ce70d3d6b22200b4508f92。

Edit the most recent generated portable radio image. Change only the speaker grille on the right side of the front face: it must contain exactly THREE horizontal rows of evenly spaced small circular holes, clearly three rows total, with no paired sub-rows and no additional holes above or below. Preserve everything else exactly: one cream-white rounded rectangular radio, one orange circular knob on the left, no screen, same body proportions, same slight three-quarter angle, same warm-gray plinth, same muted gray-green background, same soft upper-left studio light, same matte molded-plastic texture and contact shadow. Do not add text, labels, buttons, antenna, handle, cable, second radio, logo or watermark.

结果与局限：首图实际出现六排孔，一次局部修订才得到三排；不能把最终结果当首次成功。

复现边界：精确 v1 编辑输入图和实际编辑文字已收录；未归档固定 seed 或可锁定的模型构建标识，重新编辑不保证得到相同像素。

04 cyberpunk-design

已观察缺陷：unrequested third person in rear doorway。

失败首图 · 编辑输入	最终候选 · 编辑输出

打开 v1 编辑输入 → 使用下方实际编辑词 → 查看 v2 最终候选

输入图片 SHA-256：8ae96954d9f36b4d057f85f9fac2ee3993af1250cdd988996cf5463744f3395c。

Edit this underground prosthetic-clinic image by removing only the unintended background person standing in the rear doorway. Replace that figure with the same empty lit corridor and doorway surfaces. Preserve the female mechanic, seated courier, mechanical forearm and skin-liner boundary, both hand positions, repair action, cyan task lamp, tiny magenta status light, translucent divider framing, camera angle, composition, materials, lighting, colors and all other objects exactly. Keep exactly two people total. Do not add any new person, limb, object, text, logo, HUD, smoke or watermark.

结果与局限：首图多出背景人物；修订解决数量问题，但不证明义体工程可制造。

复现边界：精确 v1 编辑输入图和实际编辑文字已收录；未归档固定 seed 或可锁定的模型构建标识，重新编辑不保证得到相同像素。

07 horror-design

已观察缺陷：footprints continued beyond the empty chair。

失败首图 · 编辑输入	最终候选 · 编辑输出

打开 v1 编辑输入 → 使用下方实际编辑词 → 查看 v2 最终候选

输入图片 SHA-256：7f89f17c2318462ca458b758465d2c9b0996844ade74bf8157490e4f0ff54557。

Edit only the wet footprint evidence in this dining-room horror image. Keep the woman, dish towel, empty pulled-out chair, dining table, room architecture, closed back door, daylight, camera angle, composition, colors and all other objects unchanged. The wet bare footprints must begin at the closed back door, cross the dry wood floor, and END directly beneath the pulled-out empty chair. Remove every footprint between the chair and the foreground/woman/camera. There must be no footprints leading away from the chair and no footprint near the woman's feet. Keep the final footprint clearly visible beneath the empty chair. Do not add any person, creature, blood, text, object, fog, or new light.

结果与局限：首图脚印越过椅子；修订后终点更清楚。情绪效果仍待用户观看判断。

复现边界：精确 v1 编辑输入图和实际编辑文字已收录；未归档固定 seed 或可锁定的模型构建标识，重新编辑不保证得到相同像素。

08 noir-design

已观察缺陷：unrequested readable wall and door signage。

失败首图 · 编辑输入	最终候选 · 编辑输出

打开 v1 编辑输入 → 使用下方实际编辑词 → 查看 v2 最终候选

输入图片 SHA-256：ba9b5bddb2a0959c5959413c067786d0b32c4d50f2027a7a9579d5c337ece39a。

Edit this dockside accounting-office noir image by removing only the unintended readable environmental text: erase the wall sign text and the door words, replacing them with believable aged blank paper/panel and frosted glass texture. Keep the accountant, his expression and gaze, open ledger, burning wastebasket, wet handprint, rainy window, locked door, desk lamp, composition, lighting, colors and materials unchanged. The ledger may retain small natural accounting marks, but do not add any large readable words, captions, signage, logos or watermark.

结果与局限：修订移除墙面和门牌文字；账本自然小字仍保留。本图不能可靠证明玻璃上手印处于哪一物理表面。

复现边界：精确 v1 编辑输入图和实际编辑文字已收录；未归档固定 seed 或可锁定的模型构建标识，重新编辑不保证得到相同像素。

12 xianxia-visual-director

已观察缺陷：bright sky trails and excessive uniform detail weakened live-action photography。

失败首图 · 编辑输入	最终候选 · 编辑输出

打开 v1 编辑输入 → 使用下方实际编辑词 → 查看 v2 最终候选

输入图片 SHA-256：39681f6080bc391be14e59346560b2c3a01dbd8c94744ee2d1dd7f43b349a215。

Edit this xianxia armillary-platform image to make it read as a live-action large-format film still rather than a polished fantasy game render. Preserve the exact overall composition, human-height viewpoint, lone rear-angle adult immortal, open stone platform with no railings, left foreground rock frame, colossal armillary rings, bridge layers, floating observatories, low cloud sea and distant realm. Remove the bright magical spiral trails from the sky. Simplify the giant bronze surfaces into calmer large structural planes with restrained shallow relief, real dark patina, uneven oxidation and narrow worn edges; avoid dense equal ornament everywhere. Make the physical foreground stone slightly imperfect, damp and weighty with believable contact, and the robe visibly woven with natural folds. Reduce sharpness, microdetail, contrast and saturation progressively through middle and far distance; distant buildings should have softer silhouettes, irregular atmospheric occlusion and uneven natural illumination, while remaining readable. Use one real high-altitude morning sun from camera right, natural photographic blacks, soft highlight roll-off, subtle lens halation and fine restrained film grain. Keep the four scale proofs, five layers and 40–60% content-bearing air. No global glow, no artificial HDR, no neon edges, no cloned figures, no text, no logo, no watermark.

结果与局限：历史判定为部分通过；中远景仍有数字奇观感。公开前抽检中人物目测约20%画高，明显大于原输入约4%，精确尺度约束未通过。

复现边界：精确 v1 编辑输入图和实际编辑文字已收录；未归档固定 seed 或可锁定的模型构建标识，重新编辑不保证得到相同像素。