Repository navigation
【早期结果】在普通 DVD-R 上,让图案按设计的角度亮起来:O/X 切换、亮点合成环、红色点缀 / [Early result] Making parts of a picture light up at a chosen angle on an ordinary DVD-R: an O/X switch, spots that merge into a ring, red accents #193
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.
这张盘上同一个位置画了 O 和 X。手机在一个位置时亮的是 O,移动几厘米后换成 X。
两张都是同一张普通 DVD-R 的手机闪光灯照片,只做了透视校正,没有调亮。O 亮时比 X 亮 8 到 9 倍,X 亮时比 O 亮约 3 倍。
上面是手机慢慢移动时拍的 8 张连起来的效果:右边和左下的亮带朝相反方向走,右下两个亮点连起来再开成一个环。
最抢眼的其实是那两个亮点。下面四张是同一块地方,手机往一个方向移动:两个点,长大,连成花生,最后只剩外面一圈。
这段是转动盘时拍的。它和下面标了“调色”的照片都调过颜色,只是为了看得更清楚;文中的数字都来自没调色的原片。
这是一台光驱、一种盘、一部手机的早期结果,局限都写在后面。
和盘上本来的彩虹有什么不同
光盘本来就有彩虹,普通刻图也会出彩虹,还常常很好看。这里用的也是同样的光学现象,不是颜料。
普通刻图里,颜色主要取决于位置和角度,彩虹带从盘心散开,图案在上面挡出形状。这张盘是先定下“手机大概在哪”,再给每一小块单独挑写法,决定它在手机到哪个位置时把光送进镜头。所以亮起来的形状跟着我们画的图走,不跟着彩虹带走。
DiscImageStudio 刻的图也会出现局部和周围不同色,甚至明暗反转,这是两种写法随角度表现不同带来的。它的界面里目前没有指定颜色的选项。我们在讨论区和仓库里没找到别人按颜色设计的做法,但这只是没找到。#108 里转来的小红书视频,至今不知道是不是有意设计的配色。
一张人物盘
同样的办法也用在了一张人物图上,红色的鞋、臂章和下摆,蓝色的下摆和袖标。这张盘上不存数据。
左边是最初的设计图。中间是实际刻录的图案:灰度在盘上不够明显,所以我们把背景改成实底,卫衣提到最亮,黑裤和头发加了亮边,外圈的方块用来定位和测量。右边是我们搭的模拟器按中间这张算出的效果,用来模仿理想拍摄条件下的样子:亮度和颜色来自测试盘的实测,整张调亮了 2.5 倍方便看,红蓝点缀的颜色还是理论值。
模拟器只能看大致样子。它预测的亮暗差约 2 倍,实拍是 4 到 5 倍;顶部和底部的亮线、盘面反光也没有算进去。下面是实拍:
左边是这次的结果,右边是上一版(同一张图,存了数据,拍摄距离也不同)。卫衣和背景的亮暗差从约 1.2 倍提高到 4 到 5 倍,这是在手机接近设计位置时测的。
下面是调色后的几张,手机在不同位置,颜色和明暗都跟着变:
上面左边的动图没有调色。
做到的和没做到的都很明显:
所以说“选颜色”还太早。更准确的说法是:能决定一块地方在哪个角度亮,亮的时候大致是什么颜色。
怎么做的
一开始我们靠理论模型来挑,它预测的反差比实际高 2 到 3 倍,所以改成了先测。挑写法时用一半照片,另一半照片用来核对,避免把偶然的结果当真。
O/X 和反向流光的思路来自请 GPT 做的一份调研,那份调研是纯计算,这次是第一次实刻验证。
结果
事先写好了通过标准,再刻盘拍照:
这次盘放歪了约 13°,手机没有走到小 X 和弯道设计的位置,所以这两项还不能算设计失败,要摆正再试。
局限
很想知道有没有人试过类似的东西,或者手上的盘是不是也有这种表现。
具体实现
机位:主摄在盘心正上方 330 mm,朝图上方偏 70 mm,闪光灯相对镜头 (-43.66, +19.79) mm。这是 iPhone 17 Pro 的图纸值,用亮线方向反解,所有照片的方向误差约 1°。每张照片的机位从 ArUco 标定板解出,焦距取 EXIF。
下面的 δ 是(到灯的单位向量 + 到相机的单位向量)在轨道切向上的分量。一阶衍射满足 λ = P·δ,P 是沿轨周期。
测试盘。 32 个同心环,半径 24.6 到 55 mm,每环约 1 mm。候选字节由完整 ECC 和 EFMPlus 编码的频谱预测筛出。盘转到哪个方向都不影响,因为每环绕一圈。照片算出每环在 δ 上的线性 RGB。按字节混合的两环证实亮度约等于各字节的平均。
人物盘。 暗部全用
A5。亮部按 δ 每 0.05 一档,选两字节按 16 字节交替,例如 δ 0.30–0.35 用2D+8D,δ 0.40–0.45 用FD+50。设计时要求相邻档也不能太暗,所以窄峰的纹理(比如92)被排除了。彩色点缀用周期纹路表,单元径向 0.06 mm,切向 0.057°。效果盘。 给每一小格定一个“手机到 y mm 时亮”,算出那里的 δ,再选 P = 0.55 µm / δ 最接近的周期纹路。背景是字节
00。实测响应。 纯周期纹路的亮峰半高宽 δ 约 0.08,是我们模型(约 0.04)的两倍。峰值约为背景的 4 到 7 倍。所以白光下独立画面的通道比模型算的少,我估计约 2 个。
红绿蓝。 周期纹路扇区给出,在同一 δ 下各色用不同周期时,要等量 RGB 的面积比约为:
长周期明显更亮:同样显蓝,P = 1.76 µm 蓝通道约 0.25,P = 1.0–1.16 µm 约 0.06–0.11。这些是手机通道,不是人眼,而且 δ 0.25–0.30 解不出可靠的比例。
读回:人物盘和效果盘全盘均匀抽 96 个 ECC 块,全部一致。测试盘每环抽 3 块,96 块里 2 块读错误,没有内容不一致。
English
This disc has an O and an X drawn in the same place. With the phone in one spot the O lights up, and a few centimetres further along it's the X.
Both are phone flash photos of one ordinary DVD-R, corrected for perspective and not brightened. When the O is lit it's 8 to 9 times brighter than the X, and when the X is lit it's about 3 times brighter than the O.
That's 8 photos taken while moving the phone slowly. The bands on the right and lower left travel in opposite directions, and the two spots at lower right join and open into a ring.
The two spots are what catches the eye most. These four show the same patch as the phone moves one way: two points, growing, joining into a peanut, then only the outer ring left.
This one was taken while turning the disc. It and the photos marked as tuned below had their colours adjusted, just to make them easier to see. All the numbers here come from the untouched originals.
This is an early result from one drive, one kind of disc and one phone. The limits are listed below.
How this differs from the rainbow every disc has
Discs make rainbows anyway, and ordinary pictures often look great because of it. This is the same optics, not paint.
In an ordinary picture the colour mostly depends on where you are and the angle, with bands fanning out from the centre and the picture blocking out shapes on top. Here we start from roughly where the phone will be, and pick the writing for each small patch so that it sends light to the lens with the phone at a chosen spot. So the shape that lights up follows the drawing, not the rainbow bands.
Pictures from DiscImageStudio can also show parts that differ in colour from their surroundings, even reversing light and dark, because its two textures behave differently with angle. Its interface doesn't currently have an option for choosing a colour. We didn't find anyone else designing by colour in the discussions or the repository, but that only means we didn't find it. We still don't know whether the Xiaohongshu video shared in #108 used a deliberate palette.
A disc with figures on it
We used the same method on a picture of two figures, with red shoes, armband and hem, and blue on the other figure. This disc holds no data.
Left is the first design. In the middle is the pattern actually burned: the greys weren't strong enough on the disc, so the background became a solid dark, the hoodie went to full bright, and the black legs and hair got a bright edge. The blocks around the edge are for locating and measuring. Right is what a simulator we built works out from the middle one, imitating how it should look when photographed under ideal conditions. Its brightness and colours come from measurements of the test disc, it's brightened 2.5 times to make it easier to see, and the red and blue accents' colours are still theoretical.
The simulator only gives a rough idea. It predicted about 2 times between light and dark, where the photos reach 4 to 5 times, and it leaves out the bright lines at the top and bottom and reflections off the disc. Here are the photos:
Left is this result. Right is the previous version (same picture, with data stored, taken from a different distance). The difference between the hoodie and the background went from about 1.2 times to 4 to 5 times, measured with the phone close to the designed spot.
Here are some tuned photos with the phone in different places, where colour and brightness both change:
The animation at the top left isn't tuned.
What worked and what didn't is easy to see:
So it's too early to say we can choose colours. What we can do is decide at which angle a patch lights up, and roughly what colour it is when it does.
How it was made
We started out picking from a theoretical model, which predicted 2 to 3 times more contrast than we got, so we switched to measuring first. We picked on half of the photos and checked on the other half, so that a lucky result wouldn't count.
The O/X and opposing bands came from a study we asked GPT to do. That study was calculation only, and this is the first time they've been burned.
Results
We wrote the pass marks down before burning:
The disc sat about 13° off this time, so the phone never reached the spots the small X and the curve were designed for. Those two aren't design failures yet, and need another try with the disc straight.
Limits
If you've tried anything like this, or your discs behave this way too, I'd really like to hear about it.
Details
Pose: main lens 330 mm above the disc centre and 70 mm towards the top of the picture, flash at (-43.66, +19.79) mm from the lens. That's the iPhone 17 Pro drawing's value, and the streak direction across all photos agrees to about 1°. Each photo's pose comes from an ArUco board, with the focal length from EXIF.
δ below is the component along the track of (unit vector to the light + unit vector to the camera). The first order follows λ = P·δ, with P the period along the track.
Test disc. 32 rings from radius 24.6 to 55 mm, about 1 mm each. Candidate bytes were shortlisted from spectra of fully encoded ECC and EFMPlus blocks. Rotation doesn't matter since each ring goes all the way round. Photos give linear RGB against δ for each ring. Two rings mixing bytes confirmed that brightness is close to the average of the bytes.
Figure disc. Dark parts are all
A5. Bright parts use two bytes alternating every 16 bytes, chosen per 0.05 of δ, such as2D+8Dfor δ 0.30–0.35 andFD+50for δ 0.40–0.45. A pick had to stay bright in the neighbouring bands as well, which rules out narrow peaks like92. Accents use periodic textures. Cells are 0.06 mm radially and 0.057° along the track.Effects disc. Each cell gets "light up when the phone is at y mm". That gives δ there, and we take the periodic texture closest to P = 0.55 µm / δ. The background is byte
00.Measured response. Pure periodic textures have a peak about 0.08 wide in δ at half height, twice our model's 0.04, and reach 4 to 7 times the background. So white light offers fewer independent images than the model suggested, about 2 by my estimate.
Red, green, blue. With a different period for each colour at the same δ, the arc sector gives these area shares for equal RGB:
Longer periods are clearly brighter: for blue, P = 1.76 µm reaches about 0.25 and P = 1.0–1.16 µm about 0.06–0.11. These are phone channels, not the eye, and δ 0.25–0.30 didn't give reliable shares.
Readback: 96 ECC blocks sampled evenly across each of the figure and effects discs, all identical. On the test disc, 3 blocks per ring, 2 of 96 had read errors and none differed.
All reactions