Repository navigation
【早期结果】在普通 DVD-R 上刻出按颜色设计的图:部分花瓣显出接近原图的紫色 / [Early result] A colour-designed picture on a plain DVD-R: some petals come close to the original purple #168
Ashley-QwQ
started this conversation in
展示台 Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
平时在光盘上刻图,软件会先把彩色图片变成黑白明暗。盘上看到的颜色,是光盘自己反光带来的彩虹,不是你选的。
这张盘我们试了另一种做法:先定好颜色,再算出每一小块该刻什么。用的是普通刻录机和普通空白盘,没有改任何硬件。找到合适的角度时,部分花瓣显出了接近原图的紫色。
这是用一张盘做的早期结果,局限都写在后面。
左边是用 GPT Image 生成的原图,中间和右边是刻好的盘的照片,背景已遮掉。
颜色怎么来的
光盘背面的彩虹,来自盘上细密重复的纹路。纹路的间距不同,同一个方向上看到的颜色就不同,而这个间距是可以靠写入的数据改变的。
所以我们先定好相机和灯的位置,再给每个小格选一个间距,让红光或蓝光刚好射向相机。花瓣里的红格和蓝格混在一起,看起来就是紫色。盘上不同位置需要的间距不一样,所以每一小块都要单独算。
现在的效果
还没做到的
这张盘同时还存了 1.38 GB 数据,那是另一件事,单独写在这里:#170。
更早的讨论:颜色 #108。
颜色算法(实验性)
这是我们现在用来算纹路的方法,还在实验阶段,不保证换一张盘、一个机位也能用。
scale取原图R + 1.6 B的 99.5 分位。1.6 是折中值:少量实测里红线约是蓝线的 2 倍,模型里差不多一样强。可用的序列(周期 µm 和扰码后的字节,重复写入)。模型里谱线占比 58% 到 69%,F512 不超过 0.37%:
这个算法哪些有依据,哪些没有:
92按它预测,11 张照片 217 个观测的色相中位误差 22.2°(打乱基线 68.2°)。但这是看过数据后才定的方法。预注册的留出检验没有通过(46.8°,门槛 30°)。92以外的周期,旧数据很少。B9的颜色至今没有完全解释。技术细节
波长 = 周期 × δ,δ 是光源和相机方向的单位向量之和在轨道切向上的分量。轨道方向固定,能选的只有沿轨周期。中图的饱和度很可能是 JPEG 截断(绿色通道几乎为 0)。亮度约 0.25 对原图 0.62。
Colour algorithm (experimental)
This is how we currently work out the patterns. It's experimental, and we can't promise it holds on another disc or from another position.
scaleis the 99.5th percentile ofR + 1.6 Bin the original. 1.6 is a compromise: a little measured data has the red line about twice as strong as blue, while the model has them about equal.What this does and doesn't rest on:
92gave a median hue error of 22.2° over 217 observations in 11 photos (shuffled baseline 68.2°). But we fixed that method after seeing the data. A pre-registered held-out test did not pass (46.8°, threshold 30°).92. The colour ofB9still isn't fully explained.English
When you burn a picture onto a disc, the software normally turns a colour image into black, white and grey first. The colour you see on the disc is its own rainbow reflection, not something you chose.
On this disc we tried another way: decide the colour first, then work out what to burn for each small area. It's a normal burner and a normal blank disc, with no hardware changes. At the right angle, some petals come close to the original purple.
This is an early result from one disc. The limits are listed below.
Left: the original picture, made with GPT Image. Middle and right: photos of the burned disc, background masked (images above).
Where the colour comes from. The rainbow on a disc comes from fine repeating patterns. Different spacing shows a different colour in a given direction, and the data you write can change that spacing. So we fix where the camera and light will be, then pick a spacing for each small cell that sends red or blue light towards the camera. Red and blue cells mixed together look purple. Each spot on the disc needs its own spacing, so each small area is worked out separately.
What it looks like now.
What it can't do yet.
The same disc also holds 1.38 GB of data. That's a separate story: #170.
Earlier discussion: colour #108. The numbers are in the technical details above.
All reactions