方法与可靠性

FilmAnalysis 的每个数字都有出处：要么由程序直接从你的照片计算，要么来自一次有记录的评测。这一页说明它们怎么来、意味着什么，以及不能说明什么。

四类信息，分开看

• 直接测量：从照片像素计算的数值，例如平均明度、对比度、色彩指纹和代表色。同一张照片每次结果相同。
• 统计比较：把测量值放进参考集里比较，例如「高于参考集中 84% 的照片」。参考集是 2,243 张知识共享许可的照片（其中胶片扫描 1,743 张、数码照片 500 张），用同样的方法测量。
• 模型判断：本地模型给出的估计，例如胶片感估计和摄影师风格相似度。它们有评测数据，但不是事实。
• 文字解读：根据测量结果写出的描述，是解读，不是测量。

色彩与影调怎样测量

照片按 sRGB（D65 白点）读取。色彩指纹把每个像素按 CIELAB 色相分成 24 个扇区、按彩度分成 6 圈；代表色用 OKLab 色彩空间里的 k 均值聚类（k = 6）得到，比例合计 100%；影调按明度 L* 统计，暗部 L* < 25，亮部 L* > 75。纹理数值描述的是边缘和细节，不是胶片颗粒的实测。版本：color-study-v1。

摄影师风格匹配的可靠性

在 27 位有足够公有领域作品的摄影师上评测，测试作品不参与建库，同一组照片和近似重复不会同时出现在库里和测试里。随机挑选的命中率约 3.7%。

• 只用图像特征：第一名命中 0.50（两组测试分别为 0.50 和 0.48）。
• 网站使用的组合方法（图像特征加测量出的风格轴）：第一名命中 0.64 和 0.58。
• 去掉所有同题材作品之后再匹配：0.53 和 0.47。这说明匹配不只是在比题材，但题材仍有明显影响：图像特征最接近的作品有 58%–63% 和被比较的照片题材相同，而随机只有 16%。

相似度分数表示视觉语言的接近程度，不是「这张照片出自这位摄影师」的概率。

胶片感估计的可靠性

在 502 位从未用于训练的上传者的 1,139 张照片上评测（胶片扫描与数码照片）：区分能力 AUROC 0.916（95% 区间 0.890–0.937），校准误差 0.026；照片压缩到 JPEG 质量 50 时为 0.90。它衡量的是「像不像胶片扫描」，不能证明拍摄时用了胶片。

胶片风格相似度能说明什么

胶片风格来自 LUT 模拟。在从未用于训练的照片上，判断套用了 13 种风格中的哪一种，第一名命中 0.17–0.24（两组测试照片），随机为 0.07。它描述的是色彩和影调与某种模拟风格的接近程度。

它不能识别真实的胶片型号。在 409 位上传者的 1,146 张标注了胶片的真实扫描上，判断胶片系列的第一名命中率为 0.282，而「总是选最常见的系列」是 0.278；具体型号接近随机；也无法可靠地判断一张照片用的是不是库里没有的胶片。

评测数字怎样读

• 第一名命中率：排在第一的答案正确的比例。
• AUROC：随机取一张胶片扫描和一张数码照片，模型把胶片排在前面的概率；0.5 等于随机。
• 校准误差：模型说「70%」时，实际是否接近 70%；越小越可信。
• 这三种数字来自不同的测试，不能互相换算，也和单张照片的相似度分数不是一回事。

FilmAnalysis 不能判断什么

不能确定胶片型号、相机、镜头或后期方法；不能证明照片的作者；不能判断照片好坏。所有分析都在 FilmAnalysis 自己的服务器上完成，不调用外部 AI 服务。

返回首页上传照片 · 关于 FilmAnalysis

Methodology and reliability

Every number FilmAnalysis shows has a source: either it is computed directly from your photograph, or it comes from a recorded evaluation. This page explains where the numbers come from, what they mean and what they cannot tell you.

Four kinds of information, kept apart

• Direct measurements: values computed from the pixels, such as mean lightness, contrast, the colour fingerprint and the palette. The same photograph always gives the same result.
• Statistical comparisons: a measurement placed in a reference set, such as “above 84% of the reference photographs”. The reference set is 2,243 Creative Commons photographs (1,743 film scans, 500 digital photographs), measured the same way.
• Model estimates: estimates from local models, such as the film-likeness estimate and photographer similarity. They are evaluated, but they are not facts.
• Written interpretation: descriptions written from the measurements; interpretation, not measurement.

How colour and tone are measured

Photographs are read as sRGB (D65 white point). The colour fingerprint places every pixel in one of 24 CIELAB hue sectors and 6 chroma rings; the palette comes from k-means clustering in OKLab (k = 6) and its proportions add up to 100%; tone is measured on lightness L*, with shadows below L* 25 and highlights above L* 75. Texture values describe edges and detail, not measured film grain. Version: color-study-v1.

How reliable photographer matching is

Evaluated on 27 photographers with enough public-domain work. Test works are never in the reference library, and groups of related or near-duplicate photographs never sit on both sides. Guessing at random is right about 3.7% of the time.

• Image features alone: top-1 accuracy 0.50 and 0.48 on two different splits.
• The method the site uses (image features plus measured style axes): 0.64 and 0.58.
• With every work of the same subject removed from the library: 0.53 and 0.47. Matching is not just subject matching, but subject still pulls: the closest work by image features shares the query's subject 58%–63% of the time, against 16% by chance.

A similarity score says how close the visual language is. It is not the probability that the photographer made the photograph.

How reliable the film-likeness estimate is

Evaluated on 1,139 photographs from 502 uploaders never used in training (film scans and digital photographs): AUROC 0.916 (95% interval 0.890–0.937), calibration error 0.026; 0.90 after recompression to JPEG quality 50. It measures how much a photograph resembles a film scan; it cannot prove that film was used.

What film-style similarity can tell you

Film styles come from LUT simulations. On photographs never used in training, picking which of the 13 styles was applied is right 0.17–0.24 of the time on two test sets, against 0.07 by chance. It describes how close the colour and tone are to a simulated style.

It cannot identify a real film stock. On 1,146 labelled real scans from 409 uploaders, the best method named the film line correctly 0.282 of the time, against 0.278 for always guessing the most common line; the exact stock is close to chance; and a film that is not in the library cannot be reliably recognised as unknown.

How to read the evaluation numbers

• Top-1 accuracy: the share of cases where the first answer is right.
• AUROC: the chance that, given one film scan and one digital photograph, the model ranks the scan higher; 0.5 is guessing.
• Calibration error: whether “70%” really means about 70%; smaller is more trustworthy.
• These numbers come from different tests; they cannot be converted into one another, and none of them is the similarity score shown for a single photograph.

What FilmAnalysis cannot determine

It cannot determine the film stock, camera, lens or editing; it cannot prove who made a photograph; it does not judge whether a photograph is good. All analysis runs on FilmAnalysis's own server, without external AI services.

Back to the home page to upload a photograph · About FilmAnalysis