Repository navigation
【实验中】能在光盘数据面刻出指定颜色吗? / [Experimental] Can we burn chosen colours on the data side of a disc? #108
Replies: 21 comments 165 replies
|
This looks interesting, this weekend I hopefully will have a working dvd drive and I have already prepared 50 DVD-R's. How can I generate this test image to burn on the disc? It's not really obvious from your post. I am also interested in what can be achieved regarding the color rendering on bluray. Although I mostly consider the hacked bluray drives. I know that STM32G4 have HRTIM peripheral that can generate PWM with 184 ps resolution. What if you connect it to laser head directly? According to my calculations that gives you a diffraction grating with the ability to change its period in around 1 nm steps. This may give perfect color rendering if you align everything perfectly. |
|
@LTVA1 That would be really useful. We've never looked at the burned marks under a microscope, and right now the size and shape of the marks are the biggest unknowns in our colour model.
python -m discprobe timing --drive E: --speed 5540python -m discprobe timing --drive E: --speed 11080If one speed gives Microscope: some things that would help most, if you have the time:
How we currently think the colour works, so you can see what the photos could test. This is our working model, and parts of it are unproven. A repeating data pattern writes marks that repeat along the track, which makes a tiny diffraction grating running along the track. Its period P is set by the pattern: What we've measured so far:
This is why your microscope matters. The model needs the mark width, shape and reflectance for each run length, and we've only been able to guess them. Seeing whether One more thing to check: at low NA (say 0.2 or below), does No rush, and good luck with the drive on Saturday. |
|
@LTVA1 One more note on the microscope. We'd already thought through how to look at these discs, so here's the reasoning in case it saves you some time. None of this has been tried yet, since we don't have a microscope. Lighting matters most. A disc doesn't transmit light, so a biological microscope lit from below won't work. What you need is coaxial reflected light (epi-illumination, as on a metallurgical microscope), with the light coming down through the objective. A ring light or other oblique light sends the diffraction rainbow straight into the lens. Which kind is yours? There are two different questions, and they need different objectives:
Low NA comes first, because it's easier and answers a question we can't settle otherwise. First-order light leaves at sin θ = λ / P. With your blue LED (about 450 nm), At higher NA, one rough model of ours has For measuring anything:
If you tell me the objectives' NA once you've checked, I can suggest what to look at first and make a test ISO with labelled |
|
@LTVA1 A correction to what I told you earlier: some of those numbers are now out of date. We had assumed the phone's flash sits right next to the lens, but on this phone it's more than 4 cm away. So we'd been calculating δ wrongly. With that fixed, |
|
@LTVA1 Thanks, that helps. Levenhuk's manual lists these objectives for the 40L NG: 4x with NA 0.1, 10x with NA 0.25 and 40x with NA 0.65. Here's what I'd expect from each. None of this has been tried yet, so please treat it as a starting point.
Lighting: an LED placed beside the objective is oblique light. The mirror reflection then misses the lens, so you see only scattered and diffracted light. That's a real and useful view, similar to dark field, and it should show which areas diffract into the lens. But it answers a different question, and the rainbow may come in too. So photos taken that way can't tell us about brightness at the mirror angle. To get light straight down with no special parts, you could try a clean glass microscope slide held at about 45° between the objective and the disc, with the LED shining in from the side. The slide reflects a little light down onto the disc, and the reflection comes back up through it to the lens. It's dim, so it needs a dark room and a bright LED. It only works if the 4x has enough room below it, so please measure that gap. Is it a centimetre or two? Practical points:
If you get a useful image with the 4x, I'll make a test ISO with large labelled |
|
当前的成品实际上已经有比较相近的性质了 |
|
@LTVA1 Sorry, I need to correct myself again. I told you The rest of what I said about the objectives and lighting still stands. The top post has the details. |
|
@Catho168 谢谢,这张图和帖子都很有意思。图上人物偏暖、背景偏绿,确实很像彩色。我想确认它是不是"指定的颜色",这也是这个帖子最想搞清楚的。 我这边的猜测是:这种颜色可能主要来自导槽本身的彩虹带,图案只是在彩虹上调了明暗,所以碰巧看起来像还原了颜色。这一点我没法从一张截图确定,小红书的视频我也要登录才能看完。 有个简单的办法能分开这两种情况:把盘原地转一下,手机和灯不动。
如果你或原作者方便,想请教两件事:
|
|
@LTVA1 Thanks, I can't watch the whole video without logging in, so that helps. If the colour moves with the rainbow slice, it comes from the groove, and the picture only changes how bright that rainbow is in each spot. So that image probably isn't setting colours through the data. I'd still like a clip of the disc turning in place to be sure, but this is what we expected. |
|
@Catho168 @LTVA1 我前面请你们转盘拍一段,这不太对,那是别人发的视频,盘不在你们手里,请忽略。小红书的视频我要登录才能看完,如果有谁存了,方便的话能发一份 mp4 吗?直接拖到这个帖子里就行,我想自己看看盘动起来时颜色和图案各自怎么走。不方便发也没关系。 @LTVA1 Sorry, please ignore my request for a clip of the disc turning. It's someone else's video, so you don't have the disc. If you saved the video, could you attach it here as an mp4? I can't watch all of it without logging in, and I'd like to see how the colour and the picture each move. No problem if that isn't easy. |
|
@Catho168 这种效果确实好看,对准角度就能出彩色,也不用新纹理,这条路很实用。我们想试的是另一件事,简单说一下区别:
这也是我们还要继续做下去的原因。目前我们试的几种图案里,只有周期最长的那种(写入字节 In short: the groove rainbow gives a colour set by position and viewing angle, and the picture only changes its brightness. We're trying to make the written pattern itself choose the colour, so that two neighbouring patches can differ. So far, of the patterns we tried, only the one with the longest period (byte |
|
@LTVA1 Thank you for the video, this is exactly what I needed. Having watched it, my "groove rainbow" explanation above looks doubtful. At about 00:23, one face briefly shows red cheeks next to yellowish skin, and the groove alone wouldn't explain that. We're now checking whether this comes from the software's settings, whether the author did it on purpose, or whether it's a coincidence. I'll post what we find here. |
|
更新:我们第一次试着按目标颜色来设计一张盘,给每个位置选纹路的间距。在手动找到的角度下,部分花瓣接近原图的紫色,但只在一个角度成立,整张图也还没有全对,也不算验证了我们的计算。照片和做法在 #168。 Update: we tried designing a disc for a target colour for the first time, choosing the pattern spacing for each position. At an angle we found by hand, some petals come close to the original purple. It only holds from one angle, the whole picture isn't right yet, and it doesn't confirm our calculation. Photos and the method are in #168. |
|
之前说过看完视频会把发现发到这里,这里补上结论。主要结论:我们没能确定那段视频的颜色是怎么来的。
如果有人有原图、工程文件或做过同样的盘,欢迎发上来。我们自己按目标颜色设计的那张盘在 #168,那是另一件事,不能证明视频也是这样做的。 Closing what I promised above about the video: we could not work out how its colours were made. The narrow rainbow lines don't colour the figures, so the groove alone doesn't explain it, and I was wrong about that. From about 23.15 to 23.70 s one face does show orange-red cheeks and a yellower chin while the clothes stay green. Our "1.6 to 3.2°" estimate doesn't hold, and tracking failed in part of the clip. The disc is probably a CD, based on a second-hand report of the author's reply that we didn't see ourselves, so our DVD numbers don't apply directly. We can't tell whether the author designed the colours or a greyscale picture happened to show them at that angle. If anyone has the original picture, the project file or a similar disc, please share it. Our own colour-designed disc is in #168. That's a separate thing and doesn't show the video was made this way. |
A few notes: |
|
@Ashley-QwQ a quick question: when converting the image into iso, does discprobe lower the quality? E.g. what happens if I make a dvd iso from 16000x16000 picture? |
@Ashley-QwQ I get this error as you told me. I will try to edit the library source code to get rid of it |
|
@LTVA1 Sorry for putting you through so much troubleshooting today, and thanks a lot for your patience and all the testing! Get some good rest after this one. Hopefully we'll get to the bottom of it soon :) |
|
@LTVA1 Oh you're learning Chinese too? That's a lovely surprise haha! I'm a native Chinese speaker, by the way. :D Your schedule sounds really packed with work, university, and language classes. Please don't sacrifice your rest just to test the software! There's absolutely no rush on my end. Hope January gives you a little more breathing room. And good luck with your Chinese lessons! 加油!❤️ |
|
基本信息: 效果如图:
另附赠言: |















Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
(左边是 GPT Image 生成的原图,右边是刻好的 DVD-R 数据面的照片,背景已遮掉。)
现在的答案是部分能。在普通 DVD-R 上,找对角度时,主玫瑰的中部和花心接近原图的紫色,只是暗一些。但整张图还没全对,挪一下手机颜色就变,目前也只做过紫色这一种目标。这张图的详情在 #168,这个帖子记录整个实验的进展,包括走错的地方。
English version at the bottom.
为什么会有颜色
光盘的数据是沿轨道排列的一串长短不一的坑。反复写同一段数据,坑就会规律地重复,这一小块盘面就像一块光栅,把白光分开。重复的间距不同,同一个方向上看到的颜色就不同,而间距可以靠写入的数据来选。换个角度,颜色也会跟着变。
任何光盘都自带彩虹,空盘也有,所以要分清哪些颜色是写进去的。灯和相机都不动,看紧挨着的两块区域:自带彩虹在这两处差不多,写入的颜色可以一块绿、一块紫。
现在做到哪了
92这种纹理的色相误差中位数要不超过 30°,结果是 46.8°,没通过。之后修正了几何,不拟合时约 22°,但这是事后算的,不算通过。下一场考试还没完成B9这种纹理92在那里偏米色、橙粉。我们认为这是反射光本身的颜色,不是上面说的光栅分出的颜色,原因还没弄清想自己试
下面这种花可以自己刻。我们之前只告诉大家怎么刻测试色板,这是我们的疏漏:色板只用来测量,刻对了也不会像花。
(这是我们的第一张花盘,下面的命令生成的就是它。玫瑰盘的设计工具还没公开。)
1. 先测盘。 花要按角度摆放,所以要先知道这批盘的螺旋几何。先 clone disc-probe,刻一张写满随机数据的盘,再放回光驱测,这一步只读不写。详细步骤和光驱要求见 README。测量主要在 Windows 上做,Linux 和 macOS 的移植版也在 README 里。
2. 生成花盘,刻到同一批的空盘上。 这一步不碰光驱,任何系统都能跑,要装
numpy和pillow。--geo填你测出的summary.json。不想测盘的话,可以刻同心环。看不到花,但能看到每种纹理的差别:
python -m discprobe rings --output rings.iso。3. 用什么盘。 这张花我们只在 RITEKF1 DVD-R 上刻过。别的盘能不能显色、颜色一不一样,我们还不知道。
4. 怎么拍。 盘面像镜子,拍到的主要是它映出来的东西,所以要让它映着暗处。
5. 该看到什么。 同一朵花上有不同颜色,背景偏灰褐。不是每朵都好看,我们的照片也是挑角度拍的。如果整张发白,或者屋顶的灯映在盘上,多半是反光盖住了颜色,换个位置再拍。
常见的误解: 程序生成的预览图用颜色区分不同纹理,不是盘上会出现的颜色。盘上的颜色取决于位置和机位,所以和预览图对不上是正常的。
想请大家帮什么
B9和明暗差的问题。技术细节:纹理名字、公式、已知的限制、想请教的问题
和光盘自带的彩虹有什么不同
任何光盘,包括没刻过的空盘,都有彩虹。那是出厂压好的螺旋导槽造成的,导槽一圈圈排列,本身就是一块光栅。
最简单的分辨办法是:灯和相机都不动,看紧挨着的两块区域。自带彩虹在这两处差不多,颜色只取决于位置和角度。写入的颜色可以一块绿、一块紫,因为两块的周期不同。花瓣就是这样。
92、B9这些名字是什么名字就是我们反复写入的那个字节(十六进制)。这个字节写入前要抵消光盘的加扰,写的时候还要经过 8-16 编码,所以字节本身并不等于坑的长度。坑的长度和周期要用编码器算出来。
下表里的 T 是光盘的基本长度单位,约 0.133 µm,坑和坑之间的间隔(岸)也用它量。
92B966A5B9的 5-5-6 每 16T 重复一次,但坑和岸是交替的,上一轮的坑到下一轮变成岸,所以明暗要 32T 才真正重复。它最强的成分是 32T 的 3 次谐波,所以表里写的是 1.42 µm。后半段正好是前半段明暗反过来,所以理想情况下只有奇次谐波。颜色怎么算
目前最简单的公式是:
λ = δ × P
λ 是看到的波长,P 是上面表里的周期。δ 只和几何有关:从盘上这一点分别指向灯和相机,取这两个方向的单位向量相加,再看沿轨道方向的分量。
举个例子,δ = 0.25 时,
92对应 533 nm,是绿色。66只有 266 nm,是紫外线,看不到。δ 变到 0.6,66算出来是 636 nm(红),A5是 510 nm(绿)。这些是公式算的,还不全有实测支持,下面会讲。因为 δ 随灯、相机和位置变,所以换个角度颜色就变,一张照片里不同位置的颜色也不同。
已知的限制
想请教
B9对不上有关。discprobe/efmplus.py,测量数据在evidence/flash_lut/,新测试盘的设计在discprobe/beacondisc.py。进展记录和走错的地方(新的在上)
2026-11-06:按颜色设计的玫瑰盘,以及我们给错了命令。
新刻了一张按紫色设计的玫瑰盘,部分花瓣接近原图的紫色,详情见 #168。同一张盘还存了 1.38 GB 数据,见 #170。
ShiroiTree 想复现开头的花,照我们给的命令刻了测试色板,这本来就不会像花。这是我们的疏漏,现在“想自己试”里写的是花盘的命令。TA 的照片里,DVD+R 上的色块也显出了红色和绿色。
下面常用"色相误差"打分:把颜色放在 360° 的色轮上,预测和实测差多少度,0° 是完全一样。作为对照,我们还会把预测和色块随机打乱再算一次,这就是纯靠运气的水平。
2026-11-05(第三次更新):看了别人的视频后撤回一个判断,新测试盘也刻好了。
那段小红书视频的颜色,多半不是导槽彩虹。 有人转来一段视频,盘上的动漫人物看起来是彩色的。我们当时只听了转述,就在评论里说颜色来自导槽彩虹、图案只调明暗。LTVA1 发来录屏后,我们逐帧看了,发现这个说法多半不对:
这些都是看手持拍摄的录屏得到的,配色最好的那张脸和那一刻也是我们自己挑的。
这还不算有人实现了指定颜色。我们读了 DiscImageStudio(DIS)的代码(提交
507cac7),普通图片路径会把 RGB 压成灰度再选纹理,原图的色相没有保留。这里也更正我们之前的说法:DIS 在 CD 上有四级纹理(0D/10/63/AA),不只两种,只有两种的是 DVD 普通路径(A5/92)。原帖作者也没提到按颜色设计,写的是"光盘自带 RGB,在特定角度下还能还原图片的色彩",这句话也可能只是说明暗没有反过来。现在的猜想是,不同明暗用了不同纹理,在某个角度下碰巧落到了好看的颜色上。Catho168 说的"调色"比我们之前的说法更接近。这还没验证,要分清的话,可以刻一张灰度对照盘,把两块灰度交换,看颜色跟不跟着走。
新测试盘 colour4。 测试盘上分成很多小块(瓦片),每块放一组色块和一个编号,这样拍一张照片就知道每块是谁。旧盘的编号靠
A5和92的明暗差来读,相机一偏,大部分就读不出来了(75 块里只读出 6 到 19 块)。新盘改了这几点:92和66的差别,一半用A5和92的差别。选92和66是因为在旧照片里,就算最差的那张,这对的颜色差也最大(0.56,A5和92是 0.26,数字越大越好分)。92、66、A5、B9,理论上和B9同样的F9,纯 5T、纯 6T、纯 7T,以及和B9比例相同但排列不同的B7B2(5-5-5-5-6-6,它有偶次谐波)。盘已经用 Ritek F1 刻好。抽样读回 4001 块(每块 16 个扇区),4000 块和镜像一致,1 块读错,重读后也一致。这只说明数据写对了,颜色好不好要看拍照。
(第一张实拍。手持,手机闪光灯,加了滤镜,背景已遮住,所以颜色不能当测量用。绿框是读对编号的块,灰框和灰圈是按读对的块拟合整张盘后推算的位置。)
读取程序在合成照片上测了 9 次,没有读错编号,但一对纹理完全失效时就读不出可靠的编号了。上面这张是第一次试真盘,结果比预想的差:
(同一张盘换个角度,也加了滤镜。大部分块变暗,程序一块也认不出来,所以这张没有叠框。)
我们事先定的预期是,相机离轴约为高度的 0.31 倍时,27 块至少读对 14 块。这两张都是手持、加了滤镜,不算数,但看起来这个目标不容易达到,读取程序在考试前还要改。
(这是设计图,颜色只用来区分不同纹理,不代表盘上的实际颜色。)
试拍:相机位置怎么算,影响很大。 公式需要知道相机在哪。我们试了两种算法,一种用照片里盘的外缘,一种用垫在下面的标定纸。在旧盘上试拍 5 张,不算考试,也不拟合任何参数,
92的色相误差用盘缘是 17.5°(77 个观测),用标定纸是 61.1°(82 个观测)。两种算法差约 5 mm 和 1° 倾斜,最简单的解释是拍摄中途底座动过,所以这不能说明标定纸的方法不好。照片少,焦距也没校准,但下一场考试的主判定已经在拍摄前改用盘缘。另外,闪光灯反射点的位置和 Apple 图纸差 3 到 4 mm,这是一次不靠颜色的核对。B9又排除了一种解释。B9在预测为绿的地方实际是蓝的(见下一条)。有人提议,编码后能量可能集中在二次谐波,周期减半,颜色自然偏蓝。我们用编码器算了实际比特序列的傅里叶谱:B9最强的成分确实不在字节周期 2.13 µm,而在 1.42 µm(占 64% 的能量,不算平均亮度)。B9的 4 次谐波(1.066 µm)只有落到约 450 nm 才亮。理想情况下 4 次谐波不存在,但如果刻出来的坑和岸不对称,也许就会出现(这是推测)。按假设的光谱算,它只要有 3 次谐波的 2% 到 10%,蓝就能超过绿。但 3 次谐波那时已经偏红,所以应该呈品红。旧照片没有这段数据(δ 在 0.42 到 0.49 之间),新盘能拍到。下一步:
B9、F9和纯 5T、6T、7T 的亮带在哪里,是否出现品红。2026-11-05(第二次更新):
92不用拟合就符合公式,B9没解开,我们也收回了一句话。把闪光灯位置换成图纸值后,只用公式、图纸和盘缘算出的相机位置,没有任何拟合参数,
92在 11 张照片、217 个观测上的色相误差中位数是 22.2°,打乱后是 68.2°。按预测波长排开,实测颜色也是蓝、绿、红的顺序,超过 700 nm 就暗下去。这是看过第一次考试之后才做的,所以不算干净的考试,第一次"未通过"的判定不变。我们之前说
B9对不上"主要是 δ 算错了",这句是错的,当时的几何里有一个不合理的拟合值。用图纸几何后,B9的误差是 54.6°,打乱后是 51.0°,也就是和预测没有关系。它确实有衍射,随着 δ 增大,蓝色升到平均值的 2 到 3 倍。可是在预测为绿的 11 块上,它全是蓝的。预测为红的区域还没有数据。A5和66的观测很少,也几乎只来自一张照片,所以还不能算验证过。2026-11-05:第一次考试没通过,但发现了之前的一个大错。
我们做了一个波动光学模型,用 8 张照片拟合,留 3 张没见过的照片考试。标准事先写好:
92的色相误差中位数不超过 30°。结果是 46.8°,没通过,所以我们没有拿它去搜 RGB 方案,也没有刻盘。收获是发现之前的几何算错了:
92和A5肉眼看明暗不同,模型没有解释。长距离平均的坑占比都约 50%,但我们之前说“所以应该一样亮”说得太满了。92有两种码字,其中一种是 6T/10T,单帧岸占比可以到 37.5% 或 62.5%(模型结果,见 刻盘之后无法正常重新读取 #90),这可能参与明暗差别,也可能来自坑的实际形状或染料。要分清还得靠对照实验和显微镜。这次只有 3 张照片,彼此差别也大(
92在两张上分别是 16.6° 和 56.0°),所以这些数字都不算稳。当时的计划是固定机位补拍 36 张、再刻一张对照盘,比较"同一种图案"和"按位置换周期(P = λ / δ)"的色块哪个颜色更均匀。LTVA1 有显微镜,如果能看到坑的形状,就能回答明暗差的问题。现在第一条已改用 colour4,对照盘也排到考试之后。
2026-11-04:四色花盘和测色靶盘。 花盘同一朵花上能显 3 到 4 种颜色。测色靶盘有 75 块、18 种图案,用闪光灯拍了 11 张,得到 269 个观测。
下面是这个帖子 2026-11-06 之前的开头和状态,当时还没有玫瑰盘。
这是一张 DVD-R 的数据面,用手机闪光灯拍的。没有颜料,这些花是刻录机按我们写入的数据烧出来的。同一朵花上有 3 到 4 种颜色,换个角度颜色就变。
我们想知道能不能事先指定颜色,比如让某块区域在某个机位下显绿色。现在还做不到。这个帖子记录我们试到哪了,包括走错的地方。如果你也在试,或者懂光学、懂刻录机,欢迎指出问题。进展会更新在这一楼。
English version at the bottom.
这些花是什么
光盘的数据是沿轨道排列的一串长短不一的坑。如果反复写同一段数据,坑的长度和间距就会沿轨道规律地重复,这一小块盘面就成了一块光栅。白光照上去,往某个方向只有一种波长最强,所以看到的是某一种颜色。
每片花瓣就是一块这样的光栅,只是写入的数据不同,重复的周期也不同。这些花瓣用了 4 种数据,花心用了第 5 种。背景写的是打乱的数据,没有规律的周期,所以基本保持灰褐色。
现在做到哪了
92的颜色基本能用公式算出来,但还没通过事先定好的考试。其他图案还不行,B9尤其对不上。English
The short answer is partly. On a plain DVD-R, from the right angle, the middle and heart of the main rose in the photos at the top come close to the purple in the original, only darker. The whole picture isn't right yet, the colour changes when the phone moves, and purple is the only target we've tried. Details of that disc are in #168. This thread records the whole experiment, including where we went wrong.
Why there's colour
Data is a line of marks along the track. Write the same data over and over and the marks repeat regularly, so that patch acts as a grating. The repeat length decides which colour goes towards the camera, and you choose it by choosing the data. Change the angle and the colour changes. Every disc also has its own rainbow, so to tell them apart, keep the light and camera still and compare two neighbouring areas. The rainbow is about the same in both, while written colour can differ.
Where things stand
92and got 46.8°, so it failed. After fixing the geometry it's about 22° with no fitting, but that was after the fact and doesn't count as a pass. The next test isn't finished.B9, which is blue where we predict green. The warm colour of92near the mirror direction, which we think is the reflected light itself. The frosted look of the rose disc.Trying it yourself
You can burn the first flower disc, shown in the earlier version below. We used to give only the commands for the test target, which was our oversight: that target is for measuring and won't look like flowers even when burned correctly. The tool for the rose disc isn't published yet.
First clone disc-probe, burn a random-data disc and measure it, which only reads. The README has the steps and drive requirements. Measuring is mainly on Windows, and the README links the Linux and macOS ports.
Then build the flower image and burn it on a disc from the same batch. This step doesn't touch the drive and runs on any OS with
numpyandpillow. Give--geothesummary.jsonyou measured.We've only burned these flowers on RITEKF1 DVD-R. To photograph them, turn off the room lights and use only the phone flash. Lay the disc on something dark, hold the phone 20 to 30 cm away and 6 to 9 cm to the side, and turn it towards the centre, keeping its reflection off the disc. Move it slowly and the petal colours should change. Not every flower looks good, and our photos were taken from chosen angles. The preview image uses colours only to tell textures apart, so it won't match the disc.
How you can help
Photos from other discs and drives are useful. Please post originals without crops or filters, with the disc MID, drive model, phone model and whether the flash was on. A microscope or AFM view of the marks would help with
B9and the brightness differences. If you know optics or burners, please point out what we got wrong.Earlier version of this post: details, progress and where we went wrong
This is the data side of a DVD-R, photographed with a phone flash. There's no pigment: the burner wrote these flowers from data we chose. One flower shows 3 to 4 colours, and they change when the angle changes.
We want to know if we can choose a colour in advance, for example make an area look green from a given camera position. We can't yet. This thread records how far we've got, including where we went wrong. If you're trying this too, or you know optics or disc burners, please point out problems. Updates will go in this post.
What the flowers are
Data on a disc is a line of marks of different lengths along the track. If you write the same data over and over, the marks and the gaps repeat regularly along the track, and that patch of disc becomes a grating. Under white light, one wavelength is strongest in any given direction, so you see one colour.
Each petal is such a grating, written with different data and so with a different repeat length. These petals use 4 kinds of data and the flower centres use a fifth. The background is scrambled data with no regular repeat, and it stays roughly grey-brown.
How this differs from the disc's own rainbow
Every disc has a rainbow, even a blank one. It comes from the spiral groove pressed at the factory: the turns of the groove sit side by side and form a grating of their own.
The easiest way to tell them apart is to keep the light and camera still and look at two neighbouring areas. The rainbow looks about the same in both, because it depends only on position and angle. Written colours can be green in one and purple in the other, because the two have different periods. That's what the petals do.
What names like
92andB9meanThe name is the byte we write repeatedly, in hexadecimal. The disc's scrambling is cancelled first and the byte then goes through 8-to-16 encoding, so the byte isn't the mark length. We get mark lengths and periods from an encoder.
T below is the disc's basic unit of length, about 0.133 µm. It also measures the gaps between marks (lands).
92B966A5B9's 5-5-6 repeats every 16T, but marks and lands alternate, so a mark in one round is a land in the next and the pattern only truly repeats every 32T. Its strongest component is the third harmonic of that, which is why the table says 1.42 µm. The second half is the first half with marks and lands swapped, so ideally it has only odd harmonics.How we calculate colour
The simplest formula so far is:
λ = δ × P
λ is the wavelength you see and P is the period above. δ depends only on geometry: from a point on the disc, take the unit vectors towards the light and the camera, add them, and take the component along the track.
For example, at δ = 0.25,
92gives 533 nm, which is green.66gives only 266 nm, ultraviolet, so you can't see it. At δ = 0.6, the formula gives 636 nm (red) for66and 510 nm (green) forA5. These come from the formula, and not all of them have measurements behind them yet, as explained below.Because δ changes with the light, the camera and the position, the colour changes with the angle, and one photo shows different colours in different places.
Where things stand
92, though it hasn't passed a test set in advance. Other patterns don't work yet,B9in particular.Progress (newest first)
We often score with "hue error": put colours on a 360° colour wheel and see how many degrees apart the prediction and the measurement are. 0° means identical. For comparison, we also shuffle predictions against swatches at random and score again, which shows what luck alone would give.
2026-11-05 (third update): we take back a judgement after watching someone else's video, and the new test disc is burned.
The colour in the Xiaohongshu video probably isn't the groove rainbow. Someone shared a video of a disc with anime figures that look coloured. Going only on a description, we said in a comment that the colour came from the groove rainbow, with the picture only changing brightness. After LTVA1 sent a screen recording, we went through it frame by frame, and that explanation looks wrong:
All of this comes from a screen recording of a handheld video, and we picked the best-looking face and moment ourselves.
Nobody has shown chosen colours here yet. We read the DiscImageStudio (DIS) code (commit
507cac7): the normal image path turns RGB into grey before choosing a texture, so the original hue isn't kept. We also need to correct ourselves: DIS has four textures on CD (0D/10/63/AA), not just two. The two-texture path is DVD (A5/92). The post's author doesn't mention designing colours either, and writes that the disc "comes with RGB" and can restore the image's colours at certain angles, which may only mean the picture isn't inverted.Our current guess is that different grey levels use different textures, and at some angle these happen to land on pleasant colours. Catho168's description of it as colour grading was closer than ours. This isn't verified. A grey-level control disc could separate it: swap two grey levels and see if the colours follow.
New test disc, colour4. A test disc is divided into tiles, each with a set of swatches and an ID, so one photo tells you which tile is which. The old disc's IDs were read from the brightness difference between
A5and92, and off-axis photos could read only 6 to 19 of 75. The new disc changes this:92and66, and half useA5and92. We chose92and66because even in the worst of the old photos, that pair had the largest colour difference (0.56, against 0.26 forA5and92; higher is easier to read).92,66,A5,B9,F9(in theory the same asB9), pure 5T, pure 6T, pure 7T, andB7B2(the same 5T/6T ratio asB9in a different order, 5-5-5-5-6-6, which has even harmonics).It's burned on a Ritek F1. A read-back spot check of 4001 blocks of 16 sectors found 4000 matching the image and 1 read error, which also matched on a retry. That only shows the data was written correctly, and the colours need photos.
(First real photo: handheld, phone flash, with a filter applied and the background masked, so the colours can't be used as measurements. Green frames are tiles with a correct ID. Grey frames and circles are positions predicted from a whole-disc fit to those tiles.)
On synthetic photos the reader made no ID errors in 9 runs, though it can't give reliable IDs when one texture pair fails completely. The photo above is its first try on the real disc, and it did worse than we hoped:
(The same disc from another angle, also filtered. Most tiles go dark and the program can't recognise any of them, so this one has no overlay.)
Our expectation, set in advance, is at least 14 of 27 tiles with the camera off axis by about 0.31 times its height. These two photos are handheld and filtered, so they don't count, but that target looks hard to reach, and the reader needs more work before the test.
(Design image. The colours only separate the textures and aren't what you'd see on the disc.)
Trial shots: how we locate the camera matters a lot. The formula needs the camera position. We tried finding it from the disc's rim in the photo and from a printed board under the disc. In 5 trial photos of the old disc, which aren't a test, with no fitted parameters,
92has a hue error of 17.5° using the rim (77 observations) and 61.1° using the board (82). The two differ by about 5 mm and 1° of tilt. The simplest explanation is that the base moved during the session, so this doesn't show the board method is bad. There are few photos and the focal length isn't calibrated, but the next test's main result will use the rim, a rule set before shooting. The flash's reflection also lands within 3 to 4 mm of Apple's drawing, a check that doesn't rely on colour.One more explanation for
B9ruled out.B9looks blue where we predict green (see the next entry). Someone suggested that after encoding, the energy might sit in the second harmonic, which halves the period and would push the colour towards blue. We ran a Fourier transform on the real bit sequence from our encoder:B9's strongest component isn't at the byte period of 2.13 µm, it's at 1.42 µm (64 % of the energy, leaving out average brightness).B9's fourth harmonic (1.066 µm) only lights up near 450 nm. Ideally that harmonic doesn't exist, but it might appear if burned marks and lands aren't symmetrical (this is a guess). With an assumed spectrum, it needs just 2 to 10 % of the third harmonic's strength for blue to pass green. But by then the third harmonic has moved towards red, so the result should look magenta. The old photos have no data there (δ 0.42 to 0.49), and the new disc can reach it.Next:
B9,F9and pure 5T, 6T and 7T sit, and if magenta appears.2026-11-05 (second update):
92follows the formula with no fitting,B9is still unsolved, and we take back a line.With the flash position taken from the drawing, using only the formula, the drawing and a camera position from the disc rim, with no fitted parameters,
92has a median hue error of 22.2° over 11 photos and 217 observations, against 68.2° when shuffled. Sorted by predicted wavelength, the measured colours run blue, green, red and go dark past 700 nm. We did this after seeing the first test, so it isn't a clean test, and the first test's fail stands.We had said
B9was off "mostly because of our wrong δ". That was wrong, and it relied on an unreasonable fitted value in the geometry. With the drawing's geometry,B9has an error of 54.6° against 51.0° when shuffled, so it's unrelated to the prediction. It does diffract: blue rises to 2 to 3 times the average as δ grows. But on the 11 tiles predicted to be green, it's all blue. We have no data yet where red is predicted.A5and66have few observations, almost all from one photo, so they don't count as verified yet.2026-11-05: the first test failed, and it exposed a big mistake of ours.
We built a wave optics model, fitted it on 8 photos and tested it on 3 unseen ones. The criterion, written beforehand, was a median hue error of at most 30° for
92. The result was 46.8°, so it failed, and we didn't use the model to search for RGB patterns or burn anything.What we gained was finding that our geometry was wrong:
92andA5differ in brightness to the eye. Averaged over a long stretch both have about 50% mark coverage, but we overstated it when we said they should therefore look the same.92has two code words, and one of them is 6T/10T, with land coverage of 37.5% or 62.5% over a frame (a model result, see 刻盘之后无法正常重新读取 #90). That may play a part in the brightness difference, as may mark shape or the dye. Telling them apart will take a controlled test and a microscope.This test had only 3 photos, and they differed a lot (
92was 16.6° on one and 56.0° on another), so none of these numbers are firm.The plan then was to take 36 more photos from a fixed camera and burn a comparison disc, to see if a patch that changes its period by position (P = λ / δ) looks more even than a single pattern. LTVA1 has a microscope, and seeing the mark shapes could answer the brightness question. The first item now uses colour4, and the comparison disc comes after the test.
2026-11-04: colour rosettes and a colour target. One rosette shows 3 to 4 colours. The target has 75 tiles with 18 patterns, and 11 flash photos gave 269 observations.
Known limits
Questions
B9not matching.discprobe/efmplus.py, the measurements are inevidence/flash_lut/, and the new test disc's design is indiscprobe/beacondisc.py.All reactions