用Seedance 2.5制作1.5分钟短剧的完整流程

我有一个朴素的小型短剧想法，所以我决定用 Seedance 2.5 制作一部 1.5 分钟的电影，并分享整个过程。 首先，在没有准备任何角色、环境或其他参考的情况下，我把整个想法变成了一个故事，并将其压缩成一个 30 秒的提示词。我以 480p 生成了该版本。 这里的目标只是看看角色、灯光、构图和整体视觉方向会是什么样子，并检查提示词中是否存在任何逻辑问题或错误。 正如我之前提到的，如果你给 Seedance 2.5 的上下文超过其时长实际能处理的范围，它往往会通过加快场景速度来补偿。不出所料，第一个版本的大部分内容感觉像是被快进了。在审查了该视频后，我对故事做了一些修改。 我没有从头开始创建角色和地点，而是决定重用 Seedance 在第一个视频中已经创建的内容。 我从角色和环境清晰可见的时刻截取了屏幕截图，然后使用 Seedream 5 Pro 从中创建了角色表。对于餐厅，我从一个帧中移除了人物，并创建了一个干净的环境参考。对于充电亭，我在 Midjourney 中生成了一张图像并移除了其背景。 然后我修改了原始提示词，并将故事分成三个独立的部分，创建了三个独立的提示词。 对于前两个部分，我已经拥有的参考就足够了。对于第三部分，我意识到我还需要一个显示餐厅在夜晚结束时窗帘关闭的参考。所以我从之前的生成中抓取了另一帧，并将其作为第三个提示词的环境参考。 我在第 3 部分遇到了一个连续性问题。在最后一个镜头中，餐厅老板在餐厅前的外面行走，这破坏了连续性。 我没有重新生成整个 30 秒序列，而是用一个新的提示词只生成了那个 10 秒的特定镜头。我从原始版本中剪掉了有问题的部分，并用新的替换了它。 那时，电影基本上完成了。我添加了字幕，放大了它，并完成了剪辑。 我不是为特定的平台、出版物或比赛制作的。这主要是一个实验。 肯定有一些我可以做得不同的事情。我本可以为角色添加语音参考。我本可以为携带包的餐厅老板创建另一个参考。我还可以准备更多餐厅的角度，并将其用作环境参考。 我对餐厅的关闭窗帘版本也不太满意，所以本可以更仔细地准备一个专门的参考，而不是依赖生成视频中的一帧。 总的来说，如果你正在考虑制作一部 1 到 2 分钟的电影，首先将整个故事压缩成一个 30 秒的测试可能是一种有用的方法，看看这个想法在视觉上是否有效。 另一个有用的方法是，从你已经生成的素材中获取帧，并将它们用作下一部分的参考。这对连续性有很大帮助，而无需事先从头设计每一个资产。 包括提示词修订和生成时间，整个过程花了我几个小时。总的来说，我不得不生成八个 30 秒的片段和两个 10 秒的片段。其中一两次额外的生成主要是因为我对一些事情操之过急。 如果你对这种工作流程有任何建议、想法或技巧，我很乐意听到。 我在下面留下第一个测试视频，以及我从中提取和准备的参考，以及我第一次生成时使用的原始提示词。 [回复] Seedance 2.5 创意提示词： 一部电影般的短片跟随一位温暖的年轻韩国女服务员，在首尔一家小烤肉店拥挤的晚间服务中，直到所有人都离开后才揭示她隐藏的机械本性。她二十出头，拥有自然的皮肤质感、柔和的特征、温暖的棕色眼睛、扎起的深色头发、白色工作衬衫和炭灰色围裙。她动作迅速而自然，带着开朗真诚的微笑。在最终揭示之前，保持她的身份、制服和身体完全人类外观。 接地气的韩国社会现实主义逐渐转变为克制的科幻悬疑。温暖的钨丝灯光和橙色烤架光芒充满狭窄的餐厅；烟雾、蒸汽、霓虹灯反射和浅景深营造出触觉现实主义。服务期间使用流畅的 35 毫米手持跟踪，打烊后则采用更平静的构图、更长的保持、缓慢推进和锁定特写。在完成任务、交换眼神、关门和可见状态变化时切换。保留餐厅、厨房和后储藏室之间既定的路线。 节拍 1：在滋滋作响的桌面烤架和拥挤的顾客中开场，跟随女服务员在桌子间穿梭。她热情地迎接到达的客人，安排他们就座，接受订单，补充小菜，清理盘子，续杯，并在食客烤肉时更换烤盘；她的微笑从未消失。在自然的反应时刻，她与一对老夫妇开玩笑，立即更换孩子掉落的勺子，并隔着房间向一位常客灿烂地挥手致意。 节拍 2：建立一个反复出现的谜团。每当她转过身去，一些顾客会回头，互相耳语，轻声发笑；另一桌的两位年轻女性打量着她，并悄悄商议。女服务员只是短暂地侧目一瞥，然后继续专注于工作，没有改变她的微笑或节奏。让他们注意到的东西完全未知，并且没有证据表明他们知道她的秘密。 节拍 3：随着最后一批食客离开，通过有目的的剪辑压缩逐渐放缓的服务。女服务员擦拭桌子，对齐椅子，收集垃圾，然后在餐厅老板（一位年长女性）旁边洗碗。他们轻松的团队合作流露出多年一起工作的人之间的感情。餐厅变得空荡荡；烤架又黑又冷，一些天花板灯熄灭了，首尔的霓虹灯滑过窗户。 节拍 4：保持一个安静的中景双人镜头，老板拿起她的包转向女服务员。对话是自然的首尔韩语，口型同步准确，无字幕。老板温和地说：{오늘도 정말 고생 많았어.} 女服务员微笑着点头：{전 괜찮아요. 먼저 들어가세요. 나머지는 제가 마무리할게요.} 老板回答：{너무 늦게까지 있지 마.} 女服务员回答：{네. 내일 봬요.} 老板离开时说：{그래, 내일 보자.} 节拍 5：入口门关闭，门闩切断了街道的声音。缓慢推进女服务员，她的微笑停留了一息，然后从她的嘴和眼睛中完全消失。面无表情，她锁上入口，检查窗帘，沿着既定的路线走到后储藏室，进入并关上门。在关门时切换到储藏室内部的锁定视图，周围是清洁用品、米袋、纸板箱和墙上的插座。 节拍 6：她靠在墙上，掀起衬衫下摆露出腹部。在一个精确的特写中，她按下肚脐旁边一个类似痣或疤痕的微小标记。<一声干净的机械咔哒声。> 皮肤沿着一条完美的接缝无血地分开；一个小金属舱门打开，露出紧凑的非人类组件和小灯。她拉出一根细电缆，将其插入可见的插座，让它轻轻绷紧，然后她靠在墙上，双臂放松。 节拍 7：保持最终构图不剪切。她闭上眼睛，一动不动，直到她体内开始发出微弱的电子嗡嗡声。充电开始。她的脸上没有她整个晚上表演的温暖。她睁开眼睛；温暖的棕色虹膜慢慢变成冰冷的人造光芒，而她的面部结构保持不变。在储藏室门下，只能看到黑暗的餐厅，外面没有人。以插座安静的电力汲取结束，顾客的目光和耳语的原因仍然完全未解。 音频以层叠的韩语聊天声、滋滋作响的烤肉声、排风扇声、餐具声、脚步声和服务拟音开始，然后随着顾客离开自然变薄。老板离开后，只保留房间氛围、锁声、织物声、舱门咔哒声、电缆接触声、插座电力和充电嗡嗡声。无背景音乐。无字幕。

I had a modest little short-form drama idea in mind, so I decided to make a 1.5 minute film with Seedance 2.5 and share the whole process. First, without preparing any characters, environments or other references, I turned the entire idea into a story and squeezed it into a single 30 second prompt. I generated that version at 480p. The goal here was simply to see how the characters, lighting, composition and overall visual direction would look and to check whether there were any logical problems or mistakes in the prompt. As I mentioned before, if you give Seedance 2.5 more context than the duration can realistically handle, it tends to compensate by speeding up the scenes. As expected, most of the first version came out feeling fast-forwarded. After reviewing that video, I made a few more changes to the story. Instead of creating the characters and the location from scratch, I decided to reuse what Seedance had already created in the first video. I took screenshots from moments where the characters and environment were clearly visible, then used Seedream 5 Pro to create character sheets from them. For the restaurant, I removed the people from one of the frames and created a clean environment reference. For the charging booth, I generated an image in Midjourney and removed its background. Then I revised the original prompt and split the story into three separate parts, creating three independent prompts. The references I already had were enough for the first two parts. For the third part, realized I would also need a reference showing the restaurant at the end of the night with the curtains closed. So I grabbed another frame from the previous generation and added it as an environment reference for the third prompt. I ran into one continuity issue in Part 3. In the final shot, the restaurant owner was walking outside in front of the restaurant, which broke the continuity. Rather than regenerating the entire 30s sequence, I generated only that 10s specific shot with a new prompt. I cut the problematic section from the original version and replaced it with the new one. At that point, the film was basically done. I added subtitles, upscaled it and finished the edit. I didn't make this for a specific platform, publication or competition. It was mostly an experiment. There are definitely things I could have done differently. I could have added voice references for the characters. I could have created another reference for the restaurant owner carrying her bag. I could also have prepared more angles of the restaurant and used those as environment references. I wasn't completely happy with the closed-curtain version of the restaurant either, so that could have been prepared more carefully as a dedicated reference instead of relying on a frame from the generated video. Overall, if you're thinking about making a 1 to 2 minute film, compressing the entire story into a 30-second test first can be a useful way to see whether the idea works visually. Another useful method is taking frames from the footage you've already generated and using them as references for the next sections. It helps a lot with continuity without requiring you to design every single asset from scratch beforehand. Including prompt revisions and generation time, the whole process took me a few hours. In total, I had to generate eight 30s clips and two 10s clips. One or two of those extra generations were mostly because I rushed a couple of things. If you have any suggestions, ideas or techniques you use for this kind of workflow, I'd love to hear them. I'm leaving the first test video below, along with the references I extracted and prepared from it, and the original prompt I used for that first generation.