关于 FilmLook · 胶片感
上传一张照片,看看你的摄影语言最接近哪位经典摄影大师。
FilmLook 是什么
FilmLook 是一个读照片风格的网页工具。它做两件事:描述一张照片的视觉语言,找出作品与它最接近的几位摄影师;再单独看这张照片有多像胶片,以及差在哪里。最后可以做一张适合手机竖屏的分享卡片。
为什么做它
很多人拍了几年照片,说不清自己喜欢的到底是什么:是某种光,某种距离,还是某种颜色。滤镜和预设能让照片“看起来像胶片”,却很少解释为什么像,或者为什么不像。FilmLook 想试试,机器能不能把一张照片的视觉语言说得足够具体,让拍它的人觉得有用。
它还在测试阶段。每次分析后的评分会决定接下来改什么,所以评分对这个项目很重要。
摄影师风格匹配
FilmLook 先测量照片本身:色彩的饱和与冷暖、明暗对比、构图的重心、光线方向、颗粒和清晰度。然后由视觉模型读出那些测不出来的东西,比如情绪、和被摄对象的距离。这些读数和资料库里每位摄影师的风格档案比较,档案是根据他们公开作品整理的估计,不是测量。
比的是视觉语言,和拍了什么无关。拍一只乌鸦不会因此接近深濑昌久,拍一扇红门也不会因此接近埃格尔斯顿。结果里的相似度是一种相近程度,不是鉴定,也不是概率。没有哪位摄影师足够接近时,页面会直接告诉你。
胶片感评分
胶片感评分从 0 到 100,看的是色彩、高光、阴影和质感是否像胶片那样形成:高光是慢慢变亮还是突然变白,颜色在亮部和暗部怎样变化,颗粒是否随明暗分布。只加颗粒、暖色或发灰的暗部,拿不到高分。
分析会给出最接近的胶片风格(例如 Portra 400 或 Tri-X 400 的观感),说明照片哪里还像数码,再分成两部分建议:后期能改的,和下次拍摄才能改的。胶片名只是审美参照,FilmLook 不复刻任何胶片,也和任何胶片厂商没有关系。
为什么说胶片审美是主观的
“胶片感”不是一个物理量。不同胶片之间的差别,常常比胶片和数码之间的差别还大:Velvia 浓烈,Portra 柔和,Tri-X 粗粝。扫描仪、冲洗和后期又会把同一卷底片变成很不一样的样子。一张照片像不像胶片,取决于你心里想的是哪一种胶片。所以 FilmLook 会分开给出几项读数,而不只是一个数字,详见为什么胶片审美不能只看一个分数。
AI 看照片的局限
- 评分和匹配都是模型的判断,不是仪器测量。同一张照片再分析一次,分数可能有几分的差别。
- 模型看到的是缩小到长边 2048 像素的副本,看不到原片里最细的颗粒和噪点。
- 模型会受被摄对象影响,即使被要求只看视觉语言。觉得匹配不对时,请在评分里告诉我们。
- 摄影师资料库目前收录约 140 位摄影师,覆盖不了所有风格。
- FilmLook 不判断照片好不好。一张很美的照片可以一点也不像胶片。
隐私
照片只在你按下分析时上传,发给视觉模型服务的是去掉元数据的缩小副本。完整说明见隐私政策。
上传一张照片试试
About FilmLook
Upload a photo and discover which legendary photographer shares your visual language.
What FilmLook is
FilmLook is a web tool that reads the style of a photograph. It does two things: it describes a photo's visual language and finds the photographers whose work sits closest to it, and it looks separately at how film-like the photo is and what keeps it from being more so. You can then make a card sized for a phone screen.
Why it exists
Many people who have photographed for years find it hard to say what they are drawn to: a kind of light, a distance from people, a palette. Filters and presets can make a photo look like film, but they rarely explain why it does, or why it doesn't. FilmLook is an attempt to find out whether a machine can describe a photograph's visual language specifically enough to be useful to the person who took it.
It is still being tested. The rating after each analysis decides what changes next, so ratings matter a great deal to the project.
Photographer style matching
FilmLook first measures the photo itself: saturation and colour temperature, contrast, where the weight of the composition sits, the direction of the light, grain and sharpness. A vision model then reads what cannot be measured, such as mood and the distance from the subject. These readings are compared with a style profile for each photographer in the database. The profiles are estimates drawn from their published work, not measurements.
The comparison is about visual language, whatever the subject. A photo of a raven does not bring you closer to Masahisa Fukase, and a red door does not make you William Eggleston. The similarity shown is an affinity, not an attribution or a probability. When no photographer is close enough, the page says so.
The Film Look Score
The Film Look Score runs from 0 to 100 and asks whether colour, highlights, shadows and texture form the way they do on film: whether highlights brighten gradually or turn white at once, how colour changes from the shadows to the highlights, whether grain follows the tones. Added grain, warm tones or faded blacks alone do not earn a high score.
The analysis names the closest film look (the look of Portra 400 or Tri-X 400, for example), explains what still reads as digital, and splits its advice in two: what editing can change, and what only your next shoot can. Film names are aesthetic references. FilmLook reproduces no film and has no connection with any film maker.
Why film aesthetics are subjective
"Film look" is not a physical quantity. Films differ from each other as much as from digital: Velvia is dense and saturated, Portra soft, Tri-X gritty. Scanning, processing and editing turn the same roll into very different pictures. Whether a photo looks like film depends on which film you have in mind. That is why FilmLook reports several readings alongside the single number; see why film aesthetics cannot be reduced to one score.
The limits of an AI reading
- Scores and matches are a model's judgement, not a measurement. Analyse the same photo again and the score can move by a few points.
- The model sees a copy reduced to 2048 pixels on the long edge, so the finest grain and noise of the original are lost.
- Models are swayed by subject matter even when told to look only at visual language. If a match seems wrong, say so in your rating.
- The photographer database holds about 140 photographers and cannot cover every style.
- FilmLook does not judge whether a photo is good. A beautiful photo can look nothing like film.
Privacy
Your photo is uploaded only when you tap analyse, and the vision model service receives a smaller copy without metadata. The full details are in the privacy policy.
Try it with a photo