每天只睡5个小时、连续坚持两周后，我终于睡足了8个小时，感觉自己像重获新生一般。

After sleeping only 5 hours a day for two weeks, I finally got 8 hours of sleep, and I feel like I've been reborn. With this renewed energy, I'd like to briefly share my thoughts and the lessons I learned while making this MV. These are my experiences after burning through a huge number of credits as a complete beginner in video production. I'll keep it simple and concise, based on Seedance 2.0. First, attaching a character sheet is not optional—it's essential. Create the character sheet simply, like the attached image. If you don't have close-up images, the generation model may blur the character's facial details or interpret them however it wants. For the body and outfit, front and back views are enough. You don't need to include various facial expressions. In fact, they can even become a cause of model drift. Second, about prompt structure and conciseness. Specify the role of each image in the image reference slots. In the CHARACTER block, explicitly describe the girl's appearance in text. Number of characters appearing This turned out to be a completely unexpected variable for me. For example, if I didn't include text like "One girl, one elite guardian. Maintain visual consistency with all references.", I found that duplicate copies of each character were generated much more often than I expected. This happened especially when three or more main characters appeared in the same shot. In the STYLE block, write only the essentials about the art style, lighting, and rendering method as briefly as possible. For realistic, semi-realistic, or 2D styles that Seedance already handles well, this prompt isn't really necessary. But for 2.5D, which Seedance struggles with, it helped a little. It doesn't have nearly as much impact as image references. (Optional) Overall tempo instructions for the clip's pacing and mood. This is also optional, but in my case I needed extremely fast-paced action. Before adding this instruction, the pacing felt as slow and frustrating as a sloth. Expression lock instructions (Optional depending on your needs) Describe scene events in order, separated into CUT 1, CUT 2, and so on. Seedance follows second-by-second CUT instructions surprisingly well. Think and direct from the director's chair. Camera work Honestly, this is the most important part of video production. It's the core element that determines the mood and quality of the video. Even if the characters themselves aren't particularly attractive, compelling backgrounds combined with good camera work can create beautiful cinematic shots. I couldn't memorize all the terms, so I organized various cinematic camera movements, including angle and composition terminology, into a chart for quick reference. Final frame description—the still image at the exact moment the clip ends. This part can also be omitted, but I recommend using it if you have a specific ending frame in mind. That's it for the prompt structure. Second, keep your prompts as short and concise as possible. Especially when your scene contains too many events for the selected video length, the entire clip tends to fall apart surprisingly often. Our credits are precious, aren't they? Third, unless you specifically want it, always specify "No background music." It only gets in the way during editing. It's better to use Suno or another good external audio source. Fourth, use Mini or Fast for testing, and if your scene contains bold action, I recommend testing with Standard (Low Quality) before using those two. It's not easy to get a successful result on the first generation. Unless you've reached a professional level, this lottery called the [Generate button] will force you to revise your prompts over and over again, so saving credits is better for your mental health. Fifth, don't expect a great video from your very first Generate. It only increases your stress. Unless you're already a professional, expecting a perfect 15-second clip on your first attempt is just setting yourself up for disappointment. Try multiple times, then edit and combine the results. That's how you'll end up with a satisfying final video. Here are two clips from the many failed generations. They failed, but they still had some usable shots. Of course, I didn't use them because better ones came later, but most failed clips contain hidden gems worth saving—sometimes as much as half of the 15-second clip. The production schedule wasn't easy, and I accumulated quite a bit of fatigue, but the joy the final result gave me was greater than all of that. Unfortunately, it wasn't chosen by the algorithm, but I'd like to believe that I grew one step further through this MV. Now that I've completed the MV that started it all, I'm preparing for a new challenge. I'll probably move toward a new world-building approach while focusing mainly on the short clips that the algorithm prefers. With that, I'll end this long post, and I wish all video creator communities the very best. 👏👏

每天只睡5个小时、连续坚持两周后，我终于睡足了8个小时，感觉自己像重获新生一般。 借着这股重新焕发的精力，我想简单分享一下制作这支MV时的心得与学到的教训。 这是我在视频制作方面作为一个完全的初学者，烧掉海量积分后总结出的经验。 我会尽量简洁明了，基于 Seedance 2.0 来讲。 第一，附加角色参考图（character sheet）不是可选项，而是必不可少的。 像附图那样简单地制作角色参考图即可。 如果没有近距离特写图，生成模型可能会模糊角色的面部细节，或者随心所欲地进行诠释。 身体和服装方面，提供正面和背面视角就足够了。 不需要包含各种面部表情。 事实上，这些反而可能成为模型漂移（model drift）的原因。 第二，关于提示词的结构与简洁性。 明确指定每个图像参考槽位中图像的角色。 在 CHARACTER 区块中，用文字明确描述女孩的外貌。 出场角色数量 这一点对我来说完全是一个意料之外的变量。 举例来说，如果不写类似"One girl, one elite guardian. Maintain visual consistency with all references."的文字，我发现每个角色被生成出多个复制副本的情况比我预想的要多得多。 尤其当同一镜头中出现三个或以上主要角色时，这种情况更容易发生。 在 STYLE 区块中，只用尽可能简短的方式写出艺术风格、光照和渲染方式的核心要点。 对于 Seedance 已经处理得很好的写实、半写实或2D风格，这部分提示词其实并不是必需的。 但对于 Seedance 不太擅长的2.5D风格，它起到了一定的帮助作用。 它的影响力远不如图像参考。 （可选） 用于规定片段节奏与情绪的整体速度指令。 这同样是可选的，但就我而言，我需要极其快速的动作节奏。 在加入这条指令之前，整体节奏慢得让人抓狂，像树懒一样迟缓。 表情锁定指令 （根据需要可选） 按顺序描述场景事件，分成 CUT 1、CUT 2 等。 Seedance 对逐秒的 CUT 指令遵循得出奇地好。 以导演的视角去思考和执导。 镜头运动 老实说，这是视频制作中最重要的部分。 它是决定视频情绪与质量的核心要素。 即便角色本身并不是特别出彩，引人入胜的背景加上出色的镜头运动也能拍出极具电影感的画面。 我没办法记住所有术语，于是把包括角度与构图术语在内的各种电影级镜头运动整理成了一张图表，便于快速查阅。 最终画面描述——片段结束那一瞬间的静帧。 这部分也可以省略，但如果你心目中有一个具体的结束帧，我建议使用它。 提示词结构方面就是这些。 第二，提示词要尽可能简短精炼。 尤其是当场景包含的事件多过所选视频时长所能承载的内容时，整段片段常常会出人意料地崩掉。 我们的积分很宝贵，对吧？ 第三，除非你特别需要，否则一定要指定"No background music." 它只会给剪辑添乱。 最好使用 Suno 或其他优质的外部音频源。 第四，测试时使用 Mini 或 Fast；如果你的场景包含大幅度的动作，建议在使用这两种模式之前先用 Standard（Low Quality）进行测试。 第一次生成就想得到满意的结果并不容易。 除非你已经达到专业水准，否则这个被称为[Generate button]的抽奖会让你一遍又一遍地修改提示词，所以节省积分对你的心理健康更有益。 第五，不要指望第一次 Generate 就能得到优秀的视频。 那只会徒增你的压力。 除非你已经是专业人士，否则指望第一次尝试就得到一段完美的15秒片段，无异于给自己预设失望。 多试几次，然后剪辑并组合这些结果。 只有这样，你最终才能得到令人满意的成片。 以下是众多失败生成中的两段片段。 它们虽然失败了，但其中仍然有一些可用的镜头。 当然，我并没有使用它们，因为后来出现了更好的镜头，但大多数失败的片段都藏有值得保留的闪光点——有时甚至能占到15秒片段的一半。 制作进度并不轻松，我也积攒了不少疲惫，但最终成果带给我的快乐超过了这一切。 遗憾的是，它并未被算法选中， 但我愿意相信，通过这支 MV，我向前迈进了一步。 既然那支开启这一切的 MV 已经完成，我正在准备迎接新的挑战。 我可能会转向一种新的世界观构建方式，主要聚焦于算法偏好的短视频片段。 就说到这里吧，长文至此收尾，我祝愿所有视频创作者社区一切顺利。👏👏