VCB-Studio 团队里,有一个小组叫技术组。组员的主要任务是分析每一个项目的原盘画质优劣,决定如何处理、应当给怎样的参数。技术组也负责定期测试新的科技,比如 x264 -> x265 这些。本帖旨在为下一批技术总监培训招募候选者。目前,技术总监的缺乏是本队日常最大的瓶颈,也是长远来看最大的危机。

因为要求候选的技术总监拥有非常强的英语阅读理解能力,本文剩余部分全部用英文表述。如果不感兴趣,或者英文不够强的同学,请自行右上角,谢谢。

00:00/00:00

A member in the technical group is called a technical director, who is usually the leading and managing personnel in charge of a particular anime encoding. The workflow lays out as below:

  1. A technical director, say, called Lolicon, is assigned to a particular anime series, either by his own will or by the team;
  2. Lolicon downloads the Blu-ray source and analyzes the video, designs the pre-processing scripts and encoding parameters;
  3. Another member from the encoding group who has a very powerful machine, say, called Siscon is assigned to this series and primarily responsible for video encoding;
  4. Lolicon passes the scripts and parameters to Siscon to do the encoding, and then collects the encoded results for quality comparison and validation;
  5. Final results are then passed to be further checked, packaged and published. As the technical director of the series, Lolicon will continue to oversee the process.

So it is clear that the quality of our BDRips heavily depends on the technical directors, who must have the skills to deal with all kinds of image artifacts and design parameters accordingly. This can be extremely tedious and time-consuming since trial and error is the only approach to perfection.

Back when we started VCB-Studio, I was the only technical director. Then I started to write materials and conduct training for other group members who have strong background and interest in doing the same thing. Today, they have taken up the majority of our works. VCB-Studio would have been inactivated long ago without them. After I finally got freed from some exams, I decided now it is time to start another round of training, as more and more technical directors had retired from the team. It has became the bottleneck of our production as we only have limited resources in terms of directors who can start a new series, and if the situation goes on, eventually we will not be able do any new projects except for reseeding.

To be frank, the training is not cheap, both for the trainer and the trainees. It doesn’t mean that you have to pay or get paid for the training, but you indeed have to invest a lot of time and effort, and so do we. To make it clear: we are not conducting training for fun or for anyone’s personal hobby, rather, we are doing so for the survival of VCB-Studio. All qualified trainees will become a core member of VCB-Studio and are expected to stay active for at least 2 years.

The training will take about 150 hours, expectedly to be organized as 3-month weekend seminars. It will be conducted in a QQ group chat, where trainers and trainees can interact with each other. There will be teaching materials, assignments, and of course, exams. Along the way, we will eliminate trainees who cannot follow up, and near the end term, an “internship” will be given before the trainee is considered fully qualified for the technical group.

To apply for the training, you must:

  1. Be a university student or a Ph.D. candidate who is expected to be in school for at least another 2 years;
  2. Have some background in programming;
  3. Passionate about anime sharing;
  4. Have some knowledge about commonly used encoding tools.

You are not required to have a powerful computer with advanced CPU. However, it is still considered as an advantage.

To screen out those who “just want to try”, here we give a small assessment for all the candidates: encoding an 8-minute 720p clip.

The source video is chosen as Sword Art Online Season I, JP Blu-ray, Vol.5 00016.m2ts. It is a long-trailer of EP02 when Kirito and Asuna fight their first boss battle:

Download the source, and then find a good, existing 720p BDRip collection that has this trailer. You need to use that as a reference later.

Then, encode the source video into a 720p MKV file with the following specs:

Video: AV1, bit-rate similar (not more than 10% deviation) to your reference.
Audio: AAC, bit-rate similar (not more than 10% deviation) to your reference.

Write down where did you find the Blu-ray source (so we know you are not encoding from an secondary source), any tools/scripts you used for the encoding, and write a short paragraph of your comparison between your AV1 format rip versus the reference. The purpose is not to make a “good rip”, but a “good comparison”. Which one do you think has better quality? From which perspective? Any trade-off? And how do you make the comparison? Did you make any screenshots? so on and so forth.

Again, we don’t expect you to make a “good quality” rip for this task; you will be learning how to do that later. The purpose is to launch the very first “boss battle” in your way of the expedition, and show us your bravery and determination.

Submit your application by dropping a mail to [email protected] before 2019-02-01 with a download link. It should contain your ripped video, a document about your write-up for the rip, and your personal particulars. We are not asking you to put your real names or whatsoever, but do tell us your educational status, your background, how you know about VCB-Studio, why you are interested in applying, etc.

Be advised that seats are limited, and the task takes time to complete, so don’t wait until the last week to start. More importantly, we do NOT plan (and we CANNOT afford) to have another training after this one for the next few years to come; this is THE chance for you.

Finally, let me emphasize again: we are not conducting training for fun or for anyone’s personal hobby, rather, we are doing so for the survival of VCB-Studio.

“这不是一场闹着玩的游戏。”

——茅场晶彦,《刀剑神域》

 

LittlePox

LittlePox

VCB-Studio 前队长/压制组组长
LittlePox

作品项目列表 / Project list

Updated on 2018-12-06. (image credit: さけハラス3号) 项目注释 / Info notes T / M / L / O = 电视动画 / 剧场版 / 现场活动 / OVA或其它; UID / RS /...

阅读全文

VCB-Studio Service Pack Project / VCB-Studio 集中式补丁包计划(更新 SP9)

VCB-Studio流水线式团队分工,存在两个缺点: ——单任务效率低延迟高,和出错率较高。前者可以用多任务并行来弥补,后者则以态度来修正。 ——VCB-Studio Servic...

阅读全文

2018 年秋季组员招募 (已截止)

各位中秋节快乐。又到了 VCB-Studio 定期招募新组员的时间。 本次招新范围只涉及压制组、整理组和发布组。下面就是具体要求。 压制组 CPU: Ryzen 2700X /...

阅读全文

46 条评论

  1. VCB真的一直再压制组之中走在技术的前端,但也看得到压制的质量是操再技术总监的手上,花费的心思,时间,还有当下的状态,以及不同的人和心境都有机会影响成品,好奇一下,现在的深度学习技术这么火,也就是所谓的人工智能,有没有可能透过众人团队的力量来制作出一个完全透过人工智能决定9成9以上的原盘判断和压制参数制作,而达到人人的压制分配自动化,这么一来VCB就能无所不在,屹立不摇.

    1. 如果vcb-s的技术组,计算机科学水准,有全球Top 30 计算机学院博士研究生院的研发能力,大概是可以的。
      可惜,目前平均水准也就是国内普通高校CS系大二大三的样子。

    2. 要是有那么好的东西我是真想赶紧退休坐享其成咯……(话说这条留言一股中二风格迎面扑来,我读着都感觉好羞耻(T▽T)

      1. 从没有到有往往是最困难的过程,不过相信在许多人工智能的开发门槛不断下降下,即便是大学生水平也能VCB更上一层楼,我在计算机领域也是只菜鸡,不过目前有这样的建议,首先第一步是将目前首席技术总监的思路以尽可能完善的方式传承下来
        那么第一步便是将思路标准化,可以透过以往的蓝光源建立大量的样本库,例如不同的artifact建立成不同的库,样本完善了,就可以打造一个训练课程,例如看到了这些库的场景知道要怎么下达正确的阐述,只要通过整个训练课程就确保了基本的思路

        接着第二部曲就是比较有技术门槛的部分,将蓝光片源能够自动透过计算机的图像识别(传统手法也好AI也好)来抽取出各个存在aritifact的场景,然后就是画质的评断指数(例如透过一些锐利度等来判断基础数值)

        如此一来应该能够完成一套半自动的工具,而透过不断的优化这些,和技术的进步,希望有朝一日能达成全自动的理想

    3. 就技术上来说深度学习来选择最佳压制模式非常可行,但就目前来说,性价比不高。就两方面来说,用来训练的算法编写很复杂,就像LP大大说的,研发能力要足够(不过似乎目前已经有类似的算法了);另一方面,深度学习说白了就是大量的数据堆出来的,用来训练生成模型的话,图片级别还好,可能家用机级别电脑用上万张图片训练一星期怎么也出来了,但视频级别要更多算力和时间来训练。就这些来看,人工似乎目前更合适,长远了肯定深度学习更强。
      PS:我的导师有一个课题就是搞深度学习,很可惜,我是另一个课题搞区块链的。贵组需要区块链程序猿么?

    4. 这个其实不用DL,用grid search找hyper parameter是很传统的方法,只要有足够的算力,暴力搜索就可以了。学界的AI encoder是分析画面内容,这块还是有待研究。

  2. 还有就是。。。比起笼统的说program背景,我觉得细化到数字信号/图像处理背景更好。FFT变换这一类的东西虽然入门不难,但是要到总监那种级别最好有点feeling。。。就是在充分了解原理的情况下能够很快猜一个参数的合理区间来。。。视频编码一把成本还是挺高的(怼CPU当我没说)。。。当然这种东西经验也是可以培养的,但是要压RIP精益求精的话我觉得还是要能深入原理。。。毕竟这也算是一门艺术了。。。当然有视觉识别方面背景的人也是很好的,就像自传里写的,对于噪点的处理很可能回过头一看思路可能完全不同一样。
    不过果然还是最重要的是闲和爱,有时间和行动什么都能怼出来(吧)(QAQ)

    1. 编码器中的FFT、DCT等等基本是黑箱,不需要任何了解。
      然后编码参数到目前为止需要调的更多是psy相关,显然和FFT没关系,噪点保留与否的思路也是心理学相关。
      事实上能用到FFT的基本也就原生分辨率判断,然而判断原生分辨率基本靠经验足矣,判断不出的那一部分也不需要缩回720p。

    2. 要计算机知识背景只是为了你能看工具和滤镜的doc和以python为基础的vapoursynth脚本而已,并不要你非得学计算机图形学,反正那些玩意儿生产中到底用不用得上都是个问题

  3. 8分钟的AV1编码。。。现在AV1 codec 快到那种程度了???
    “使用AV1默认的速度预设,在48核HP Z840工作站上进行了5秒钟的Tears of Steel 1080p源素材编码需要23小时46分钟(1,426分钟)”
    来源: https://blog.csdn.net/vn9PLgZvnPs1522s82g/article/details/82935174 2018/10/03
    就算假设这位哥用的是60fps源,不严谨的算一下默认预设编一把大概要1426*(480/5)*(24/60)*(1/2.25)=24337min (405小时/16天)
    就算两个月的时间来试错我觉得也是时间堪忧啊。。。。(或许这就是要成为技术总监应该用普通处理器跨过的难关吧【大雾】【逃】)
    当然我只有双核i5,av1试都没试过,实际如何我是不知道的。。。我就是单纯想吐槽一下”You are not required to have a powerful computer with advanced CPU. However, it is still considered as an advantage.“

    1. 你算的很有道理,只不过我们组内报名的同学已经压完第一版了。码率没控制好,回炉重做中。另外他不是压制组的,他负责分流。

    2. 1、我不知道所谓Z840工作站主频多少,但是通过前期测试下来,AV1编码器多核优化几乎没有,48核和4核就没有区别了。
      2、VCB-S分流组的CPU战斗力本身就很强

欢迎留言

This site uses Akismet to reduce spam. Learn how your comment data is processed.