Nano Banana 2 Lite vs Image 2.5：移軸微縮硬對照

提示詞

INTENT     selective attention made geometric; miniaturisation, or a slice of focus cutting across the world
UNKNOWNS   tilt angle ;  hinge distance J = f / sin(tilt) ;  aperture N ;  focus distance
LAWS
  Scheimpflug   subject plane, lens plane and image plane extended meet in ONE common line
                ⇒ tilting the lens TILTS the plane of sharp focus in the world
  hinge rule    the focal plane pivots about a hinge line parallel to the lens plane at J below it
                ⇒ the in-focus volume is a WEDGE: zero thickness at the hinge, widening with distance.
                  Near sharp band NARROW, far sharp band WIDE — a lawful, monotone flare.
  the inversion  sharpness is governed by distance FROM THE TILTED PLANE, not by depth.
                ⇒ objects at very different camera distances can be equally sharp if they lie on the plane
                ⇒ objects at the SAME camera distance can differ in sharpness if one sits off it
                (⇒ the "equal depth ⇒ equal blur" FORBIDDEN of the thin-lens prompt is here REVERSED —
                  which is exactly why fake tilt-shift is detectable)
  vertical objects  a lamppost, tower or figure standing across the plane is sharp only in the BAND where
                it intersects, blurring above and below — but that band's HEIGHT varies with its distance
  shift         rise/fall keeps sensor parallel to the façade ⇒ verticals stay PARALLEL while framing high;
                the horizon sits off-centre. Cropping a tilted-up frame instead leaves converging verticals.
  aperture      stopping down widens the wedge; the bokeh still obeys one aperture, one blade count
OVERDETERMINATION   the plane recovered from ≥3 sharp features at differing depths — one plane fits all;
                    the wedge's widening rate must agree with the recovered tilt and N
HONEST      one tilted plane; sharpness monotone in distance from it; the sharp band widens with depth
FORBIDDEN   a blur gradient that is a function of IMAGE HEIGHT rather than of distance from a world plane
            (the post-filter tell: an even soft band top and bottom, ignoring the scene's geometry) ;
            a sharp band of constant width from foreground to horizon (no wedge) ;
            two objects on the same recovered plane, one sharp and one soft ;
            a tall vertical uniformly sharp along its whole height while the plane clearly crosses it ;
            "shifted" architecture with converging verticals, or unshifted verticals with a centred horizon
            that the framing contradicts ;
            miniature-faking by blur and saturation alone with a ground-level viewpoint no model would have.

中文翻譯

INTENT     把選擇性注意力幾何化；微縮化，或一道橫切世界的清晰焦點切片
UNKNOWNS   傾角 ;  鉸鏈距離 J = f / sin(tilt) ;  光圈 N ;  對焦距離
LAWS
  Scheimpflug   被攝物平面、鏡頭平面與像平面延長後交於同一條共同直線
                ⇒ 傾斜鏡頭會使世界中清晰對焦的平面隨之傾斜
  hinge rule    焦平面圍繞一條鉸鏈線旋轉，該線與鏡頭平面平行、位於其下方 J 處
                ⇒ 合焦體積是一個楔形：在鉸鏈處厚度爲零，隨距離增大而變寬。
                  近處清晰帶窄、遠處清晰帶寬 —— 一種有規律的、單調的擴張。
  the inversion  清晰度由到傾斜平面的距離決定，而不是由景深決定。
                ⇒ 相機距離相差懸殊的物體，只要位於該平面上，就可以同樣清晰
                ⇒ 相機距離相同的物體，若其中一個偏離該平面，清晰度就可以不同
                （⇒ 薄透鏡 prompt 中「相同景深 ⇒ 相同虛化」的 FORBIDDEN 在此被反轉 ——
                  而這正是僞造的移軸效果能夠被識破的原因）
  vertical objects  橫跨該平面的路燈、塔或人物，只在與其相交的那條帶上清晰，
                上下兩側則變得模糊 —— 但那條帶的高度隨其距離而變化
  shift         升降（rise/fall）使傳感器保持與建築立面平行 ⇒ 高位取景時垂直線仍然保持平行；
                地平線不居中。若改爲裁剪一張上仰的構圖，垂直線則會匯聚。
  aperture      收小光圈會加寬楔形；虛化仍然只服從一個光圈、一個葉片數
OVERDETERMINATION   從不同景深處至少 3 個清晰特徵反推出的平面 —— 一個平面即可擬合全部；
                    楔形的加寬速率必須與反推出的傾角和 N 相符
HONEST     一個傾斜平面；清晰度對到它的距離單調；清晰帶隨景深增大而變寬
FORBIDDEN   由圖像高度、而非到世界平面的距離所決定的模糊漸變
            （後期濾鏡的破綻：上下各一條均勻的柔和帶，無視場景的幾何） ;
            從前景到地平線寬度恆定的清晰帶（沒有楔形） ;
            位於同一反推平面上的兩個物體，一個清晰、一個模糊 ;
            一個高聳的垂直物體沿其整高均勻清晰，而該平面明顯穿過它 ;
            「已移軸」的建築卻出現垂直線匯聚，或未移軸的垂直線配上居中的地平線
            且構圖自相矛盾 ;
            僅靠模糊與飽和度僞造微縮感，卻採用任何模型都不會有的貼地視點。

溯源

原作者

Gdgtify

發佈日期

2026-09-16

原始出處

查看原帖 ↗

收錄來源

xianyu110/awesome-gpt-image2.5 · 2100356053095641424 · MIT