王家卫风格提示词产品 · 文本评审

在售提示词产品「Three Hong Kong Nights Film Still」(PromptBase,$6.99)· 本页供外部模型审阅提示词文本并给整改建议 · 非盲审

给审阅模型的说明


你是一名资深 AI 提示词工程师,同时是电影视觉顾问。下面是一款在售的图片生成提示词产品(卖给图片生成模型的用户使用),目标是:稳定产出王家卫三部影片——《重庆森林》《堕落天使》《花样年华》——各自的视觉系统。

这款产品刚用生成图完成了一轮外部盲审,结论附在本页「盲审结论摘要」区。本页不是盲审:你知道全部背景,要的就是针对性意见。

请针对提示词文本给出可以直接落地的整改建议:
- 不要只给原则性评论(例如「应该减少 AI 味」这种话不算答案);
- 每个问题都要给出改写后的具体词句(英文提示词片段),让我们能直接替换进模板;
- 如果你认为某处原文已经写对了,也请明说,不要为改而改。

作答顺序:先读盲审结论摘要,再读提示词全文,最后逐项回答页面底部的 Q1–Q6。

盲审结论摘要(刚完成的一轮图片盲审)

审阅对象:用本页提示词生成的 11 张图(模式示例 4 张+标志场景 S1–S7 共 7 张),由另一名外部审阅者在不知道图号对应关系的情况下盲审。本轮审阅方在审前看到过网址中的缩写,归属判定可能受此影响,下方分数按其自述「按画面独立评判」记录。

总体成绩:导演归属 11/11 全部命中王家卫;逐图的影片归属 11/11 全对;平均风格准确度 3.7/5;11 张中 8 张可冒充真实电影剧照;AI 痕迹判为「严重」的 0 张(轻微 6 张、几乎看不出 2 张、明显 3 张)。

3 张被判「不能冒充真实剧照」的图及逐项死因:

共同死因:三张全栽在小文字/标识类细节上(证件字、肩章、日历字形、钟面),画面大感觉均无问题。

核心批评(审阅方原文大意):「这组图最大的问题不是生成技术,而是对王家卫的模仿主要依靠高辨识度的视觉符号——金发墨镜、凤梨罐头、警察、鱼眼镜头、旗袍、饭盒。这些元素大量集中出现,容易给人复刻经典剧照的感觉。真正还没被验证的是:不依赖经典道具时,是否仍能保持相同的导演风格;以及人物与环境之间的自然关系、不同景别与照明条件下的一致性、运动模糊与细小文字与手部结构的可信度。」

另一处点名:图 01(S3 公寓女孩,鱼缸+玩具飞机那张)被评「人物、鱼缸和玩具飞机的组合更像精心布置的摄影作品」,布置感/摆拍感偏强(该图准确度 3/5,AI 痕迹判为几乎看不出,但风格可信度被布置感拉低)。

分模式平均分:堕落天使系 4.0、花样年华系 4.0、重庆森林系 3.5(被上述文字细节图拖低)。

提示词全文(产品原文,未作任何删改)

========================================================
PRODUCT 1 — Three Hong Kong Nights (v2, rebuilt from frame-by-frame
analysis of the actual films — see research2.md)
========================================================

WHY V2: The three films this product is named after are NOT one look.
Measured frame by frame, they are three different visual systems:
amber tungsten bars (Chungking), total fluorescent-green invasion
with wide-angle distortion (Fallen Angels), and lamp-lit crimson
restraint with zero neon (In the Mood for Love). This template ships
one master line plus three mode blocks with genuinely different
light sources, color placement, halo shapes and lenses.

PRODUCT CONTENTS: 1 CORE line + 7 mode blocks (A / A-DAY /
A-MOTION / B / B-RED / B-CYAN / C, variables left open) + 7
SIGNATURE SCENES presets (S1-S7, casting, props and situation
filled in — the instantly recognizable frames: blonde wig and
sunglasses, the policeman at the midnight snack bar, the girl
dancing in the apartment, the fisheye face in the green tunnel,
the jukebox in the red room, the near-touch passing in the alley,
the pineapple can and the clock).

SUGGESTED TITLE (limit 40):
Three Hong Kong Nights Film Still
Fallback: Wong Kar-Wai Style Film Still

SUGGESTED PRICE: $6.99 (unchanged from v1 research)

DESCRIPTION DRAFT (limit 500 chars):
"One prompt, three real Hong Kong cinema systems — not generic neon.
MODE A: amber tungsten bar nights, half the frame crushed to black,
red only as tiny accents, plus step-printed motion smear and milky
fluorescent daytime variants. MODE B: fluorescent green that floods
even skin tones, light tubes burning into long streaks, extreme
wide-angle faces. MODE C: 1962 lamp light, crimson and olive, a black
foreground band across the frame, zero neon. Built from frame-by-
frame analysis of the original films."

--------------------------------------------------------
MASTER TEMPLATE (pick ONE mode block, paste it after the core)
--------------------------------------------------------
CORE:
Cinematic film still of [SUBJECT] [ACTION], in [LOCATION], Hong Kong
[YEAR], shot on 35mm film, heavy photochemical grain, soft focus,
photorealistic, no text, no watermark

MODE A — CHUNGKING EXPRESS (amber bar night):
lit only by low tungsten bar lamps, the whole frame dyed amber-
orange, almost half the frame crushed to near-black brown shadow,
skin fully amber-toned, red appearing only as tiny accents (lips, a
tie, a glint on a glass), tiny specular glints on glass and metal
edges, one out-of-focus foreground figure blurred at the frame edge,
lonely 3 a.m. stillness. No neon signs, no large light blooms, no
green on skin.
MODE A-DAY (milky fluorescent variant): harsh fluorescent tube
light, overexposed milky whites, pale cyan-green tint on walls,
highlights blown out on shirts and faces, low saturation, flat
candid documentary feel.
MODE A-MOTION (step-print variant): the whole frame step-printed,
every moving figure including the subject's own arms and torso
smeared into horizontal streaks of color, only the subject's face
momentarily readable, bright daylight, desaturated street colors.

MODE B — FALLEN ANGELS (fluorescent green system):
lit by fluorescent tubes, the ENTIRE image flooded with green —
skin tones, shadows and air all green — a ceiling light strip
burning into one long white-green streak across the frame, extreme
wide-angle lens, the subject's face distorted and huge in the
foreground, deep murky green-black shadows (40-65% of the frame),
restless handheld energy. Skin is NOT protected from the green.
VARIANT B-RED: deep red walls as the environment color, green-white
fluorescent tubes as the light source falling on a pale warm face.
VARIANT B-CYAN: bright cyan-teal interior, one huge out-of-focus
white foreground mass occluding about a third of the frame, subject
thrust toward the camera.

MODE C — IN THE MOOD FOR LOVE (lamp-lit 1962, zero neon):
1962, lit only by shaded table lamps, wall lamps and a single bare
bulb with one soft round halo, deep crimson and amber tones with
olive-green painted walls, saturated color living only in the
fabric of a floral cheongsam, a huge out-of-focus black foreground
band (doorframe, furniture, a passing shoulder) occluding 25-40%
of the frame, faces in profile with lowered eyes, half lost in
warm brown shadow, telephoto compression, still, symmetrical,
restrained. ABSOLUTELY no neon, no green-tinted skin, no motion
blur, no city lights.

--------------------------------------------------------
SIGNATURE SCENES (S1-S7 — ready-to-paste presets)
--------------------------------------------------------
The mode blocks above carry the light and color systems; these
presets carry the IMAGE SIGNATURES — the specific castings, props
and situations a viewer recognizes in one second. Each preset is
a complete prompt (CORE with variables filled + its bound mode
block). For the closest match, attach the noted original frame
as a style reference and keep the preset unchanged. Verified
outputs: wkw/signature-1.webp ... signature-7.webp.

S1 — BLONDE WIG AT THE MANSION GATE (binds MODE A)
情境:《重庆森林》第一个故事——金发假发+墨镜+雨衣的女人,
夜里站在重庆大厦式旧楼入口的人流里,纹丝不动。
Ref frame: ref-wkw/chungking-1-bar-amber.jpg
"Cinematic film still of a woman in a short curly blonde wig,
dark sunglasses and a transparent raincoat, standing still at
night at the crowded entrance of an old Hong Kong tenement
block like Chungking Mansions, streams of blurred passersby
flowing around her, amber street lamps glowing through humid
night air, Hong Kong 1994, shot on 35mm film, heavy
photochemical grain, soft focus, photorealistic, no text, no
watermark. Lit only by low tungsten lamps, the whole frame
dyed amber-orange, almost half the frame crushed to near-black
brown shadow, skin fully amber-toned, red appearing only as
tiny accents, tiny specular glints on glass and metal edges,
one out-of-focus foreground figure blurred at the frame edge,
lonely 3 a.m. stillness. No neon signs, no large light blooms,
no green on skin."

S2 — THE POLICEMAN AND THE CHEF SALAD (binds MODE A-MOTION)
情境:《重庆森林》663 号警察——独坐午夜快餐店柜台,面前
一份厨师沙拉,身边所有人都在拖影里流动,只有他静止。
Ref frame: ref-wkw/chungking-3-snackbar-phone.jpg
"Cinematic film still of a young Hong Kong policeman in
uniform sitting alone at the counter of a bright fluorescent-
lit midnight fast-food snack bar, a chef salad in a plastic
container in front of him, gazing into the distance, Hong Kong
1994, shot on 35mm film, heavy photochemical grain, soft
focus, photorealistic, no text, no watermark. The whole frame
step-printed: the customers around the counter are smeared
into horizontal streaks of motion, while the policeman sits
almost perfectly still, his face momentarily readable and calm
amid the blur, bright desaturated fluorescent interior colors,
milky blown highlights."

S3 — DANCING IN SOMEONE ELSE'S APARTMENT (binds MODE A-DAY)
情境:《重庆森林》阿菲——穿着宽大男式衬衫在阳光小公寓里
随《California Dreamin'》晃动,金鱼缸和玩具飞机在旁。
Ref frame: ref-wkw/chungking-3-snackbar-phone.jpg
"Cinematic film still of a young woman with a short ear-length
bob haircut wearing an oversized men's white shirt, swaying
and dancing alone to music in a small sunlit Hong Kong
apartment, eyes half closed, lost in the song, a glass
goldfish bowl with goldfish and a toy airplane clearly visible
in the room, Hong Kong 1994, shot on 35mm film, heavy
photochemical grain, soft focus, photorealistic, no text, no
watermark. Harsh window and fluorescent light, overexposed
milky whites, pale cyan-green tint on the walls, highlights
blown out on her shirt and face, low saturation, flat candid
documentary feel."

S4 — FISHEYE FACE IN THE GREEN TUNNEL (binds MODE B)
情境:《堕落天使》——超广角鱼眼怼脸,整条隧道和整张脸
一起被荧光绿吞掉。
Ref frame: ref-wkw/fallenangels-1-tunnel.jpg
"Cinematic film still of a woman's face shot with an extreme
wide-angle fisheye lens, her face huge and distorted, pressed
close to the camera, inside a green fluorescent-lit pedestrian
tunnel, Hong Kong 1994, shot on 35mm film, heavy
photochemical grain, soft focus, photorealistic, no text, no
watermark. Lit by fluorescent tubes, the ENTIRE image flooded
with green — skin tones, shadows and air all green — a ceiling
light strip burning into one long white-green streak across
the frame, deep murky green-black shadows, restless handheld
energy. Skin is NOT protected from the green."

S5 — THE JUKEBOX IN THE RED ROOM (binds VARIANT B-RED)
情境:《堕落天使》——深红房间里,女人挨着老式点唱机抽烟,
绿白荧光管的光落在她苍白的脸上。
Ref frame: ref-wkw/fallenangels-3-red-corridor.jpg
"Cinematic film still of a woman leaning against a vintage
jukebox in a deep red room, smoking a cigarette, eyes
lowered, Hong Kong 1994, shot on 35mm film, heavy
photochemical grain, soft focus, photorealistic, no text, no
watermark. Deep red walls as the environment color,
green-white fluorescent tubes as the light source, their pale
green-white light falling on her pale warm face, red lips the
only saturated accent, murky brown-red shadows, lonely
late-night stillness."

S6 — PASSING IN THE ALLEY, ALMOST TOUCHING (binds MODE C)
情境:《花样年华》——雨夜窄巷,提着饭盒的旗袍女人与西装
男人贴身错过,肩膀将触未触,头顶只有一盏灯。
Ref frame: ref-wkw/moodforlove-2-alley-dark.webp
"Cinematic film still of a woman in a floral cheongsam
carrying a stacked metal lunch tin passing a man in a dark
suit in a narrow rain-wet alley at night, their shoulders
almost touching but not touching, both looking down, a single
street lamp glowing above them, Hong Kong 1962, shot on 35mm
film, heavy photochemical grain, soft focus, photorealistic,
no text, no watermark. Lit only by shaded lamps and a single
bare bulb with one soft round halo, deep crimson and amber
tones with olive-green painted walls, saturated color living
only in the fabric of the floral cheongsam, a huge out-of-
focus black foreground band occluding a third of the frame,
faces in profile with lowered eyes, half lost in warm brown
shadow, telephoto compression, still, restrained. ABSOLUTELY
no neon, no green-tinted skin, no motion blur, no city
lights."

S7 — THE PINEAPPLE CAN AND THE CLOCK (binds MODE A, still life)
情境:《重庆森林》时间母题静物——凤梨罐头与钟/日历在琥珀
光里同框;罐头标签必须虚化不可读,不得出现大字。
Ref frame: ref-wkw/chungking-2-bar-closeup.jpg
"Cinematic film still, a still life close-up in warm amber
lamplight: a single unopened can of pineapple chunks standing
beside an old round wall clock and a paper calendar page on a
small kitchen table at night, dust motes in the lamplight,
deep brown shadows around them, no people, Hong Kong 1994,
shot on 35mm film, heavy photochemical grain, soft focus,
photorealistic. Lit only by low tungsten lamps, the whole
frame dyed amber-orange, almost half the frame crushed to
near-black brown shadow, red appearing only as tiny accents,
tiny specular glints on the metal can and the glass clock
face. Any printed label on the can is blurred, out of focus
and completely illegible — no readable words, no large text
anywhere."

--------------------------------------------------------
VARIABLES
--------------------------------------------------------
[SUBJECT] who is in the frame (a woman in a blonde wig and
sunglasses / an off-duty policeman / a landlady carrying a lunch
tin / two strangers passing on a staircase)
[ACTION] what they are doing, kept small and quiet (resting her
chin on her hand / running through a market crowd / leaning asleep
on a shoulder / passing without touching)
[LOCATION] bar counter, snack bar, underpass, red-walled corridor,
staircase landing, restaurant booth, night street
[YEAR] 1994 for Modes A/B, 1962 for Mode C

--------------------------------------------------------
MODEL NOTES
--------------------------------------------------------
- Works text-only on FLUX-class and ChatGPT Image-class models.
- For the closest match to the examples, attach any still from the
  matching film as a style reference and keep the mode block
  unchanged — the mode blocks were calibrated against the original
  frames.
- All 4 listing examples are unedited model outputs generated with
  this template (Modes A, A-MOTION, B and C respectively).
- Common failure: models add neon signs to Mode C or protect skin
  from green in Mode B. Re-roll, or strengthen the mode block's
  last sentence.

请逐项回答的问题

Q1 逐个模式块点评。模板共有 7 个模式块(MODE A、A-DAY、A-MOTION、B、B-RED、B-CYAN、C)。请逐块点评:哪个块写得准、哪个块有问题(词义含混、互相矛盾、模型容易误读、与原片不符等),并对有问题的块给出改写后的完整词句。

Q2 小文字/标识穿帮怎么从提示词层面整改。盲审 3 张失败图全死在小文字与标识(证件字、肩章、日历字形、钟面)。请给出你的方案:是加负面词、改构图从源头避开这类元素、还是换一种表述让模型把它们处理成不可读状态?请针对证件、制服肩章、日历纸面、钟面四种情况分别给出可直接粘贴的词句,并说明每种情况你选哪条路线及原因。

Q3 不依赖标志道具,如何只靠光色/镜头/质感保持辨识度。审阅方指出本产品的辨识度主要靠经典道具符号。请针对每个模式(至少 A、B、C 三个主模式),给出 1–2 句可加入模式块的强化词句——只用光源、色温与色彩分布、镜头与景深、颗粒与质感、人物与空间关系等语言,不许出现新的标志性道具。

Q4 哪些现有词反而会加重 AI 味。请通读全文,指出哪些词或表述容易让画面变得过度干净、过度锐利、或带来摆拍/布置感(例如 S3 被评布置感强),建议删掉或改写哪些,给出替换词句。没有把握的不要硬指。

Q5 结构与变量设计有无改进。现结构为:1 条 CORE 主干+7 个模式块(粘贴在主干后)+7 条 SIGNATURE SCENES 预设;变量为 [SUBJECT] [ACTION] [LOCATION] [YEAR]。请评价这套分工是否合理:变量设置是否够用或多余、模式块与预设的职责是否重叠、有没有更稳的组织方式。给出你的改法(如需改结构,请给出改后的结构示例)。

Q6 其他建议。任何能提升出图风格准确度与真实感的建议都可以写在这里:包括但不限于负面提示词的写法、参考图的使用方式、针对不同模型(FLUX 类、ChatGPT Image 类)的适配差异等。请同样以可落地的词句或步骤形式给出。

作答要求重申:每问都要给可直接替换进模板的英文词句,不要只给原则性评论。谢谢审阅。