Wan26多镜头

Multi-shot video in one prompt is mostly a formatting problem. The model needs to know which shot happens when, how the camera behaves in each, and how they join - and writing that by hand in the order Wan likes is where takes go wrong. This node writes it for you.

What it is, and where it fits

Wan26多镜头 is a prompt-assembly node for Wan 2.6, the December 2025 release that added shot switching and character interaction - and which, like 2.5 before it, shipped commercial-only. No open weights, so this feeds the Wan 2.6 API node, not a local checkpoint. The author's own note is that the 2.5 node works fine as a plain image-to-video node on 2.6 too; this is the one you want when the clip is a sequence.

The reason this is worth a node rather than a paragraph: Wan responds well to time-stamped shot blocks, and multi-shot prompting has been a known (if fiddly) technique on the open 2.2 weights for a while - people post consistent scene and character results from prompt structure alone. This formalises the structure and numbers the shots.

How it builds the prompt

You get three outputs. 多镜头提示词 is the shot list itself: an optional overall-description sentence, one transition sentence, then each shot as a Chinese ordinal, a bracketed time range and the description:

这个故事以第三人称视角，讲述了一个关于放弃与重拾希望的短剧。
镜头间使用硬切转场。
第一个镜头[0-3秒]一个男孩在操场的角落独自坐着，低头望着手中的信纸…，镜头推近主体
第二个镜头[4-6秒]硬切转场，固定机位，聚焦于男孩的眼睛…

完整提示词 wraps that with a total-duration line up front and your overall description and extra instructions labelled at the end - the one to actually feed the API node. 镜头结构表 is a plain-text table (序号 | 时间范围 | 时长 | 内容概要) of what you just built. It's a readout for you, not a prompt; wiring it into anything is a mistake.

Camera movement only gets appended to a shot if you pick something other than 无, and it's trimmed down to the core movement clause rather than the full speed/duration sentence the I2V node uses.

The inputs that matter

Three required, and the third one is a trap door:

• 总体描述 - the story summary. Worth writing properly: it's the only place the model learns that shot 3 relates to shot 1.
• 镜头数量 - a 1–5 slider.
• 启用智能多镜 - a boolean. Turn it on and the node throws away every per-shot field and instead emits a single prompt asking Wan to build a 3–5 shot narrative itself from the overall description, with your 附加指令 appended if you wrote one. The shot-list output and the complete prompt become the same string, and 镜头结构表 becomes the literal text 智能多镜模式已启用. Nice when you have no shot breakdown; useless when you do, and it silently ignores everything you typed per shot.

The per-shot blocks are all optional: 镜头开始时间 / 结束时间 as text (0秒, 3秒), 运镜方式 (the same 35 movements as elsewhere in the pack), and 镜头描述. Plus 转场效果 (硬切转场 default, 淡入淡出, 交叉溶解, 滑动转场, 无), 时间格式 (秒 or 帧), 视频总时长, and 附加指令 - whose default is a sentence about keeping subject, scene and mood consistent and the narrative coherent between shots. That default is good; keep it unless you're deliberately cutting loose.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/a63976659/ComfyUI-prompt-formula.git

Or use ComfyUI Manager and search prompt-formula. Restart, then it's under 📕提示词公式 / 图转视频. Nothing to download - the pack needs no models and requirements.txt is a single torch>=2.0.0 line.

Gotchas

Raising 镜头数量 alone doesn't add shots. Shots 4 and 5 are only included when the slider is high enough and their description field is non-empty. Set the slider to 5, leave 镜头4描述 and 镜头5描述 blank, and you get a three-shot prompt with no warning. Fill the descriptions first, then the slider is just what makes them visible.

The slider, in other words, is a UI control - not a count. If you're on an older version of the pack, this is the same mechanism the author described when the shot count was moved onto a slider: one glance tells you the shot count, but the descriptions are what actually drive the output.

帧 mode is 30fps, and it also relabels the total. Switch 时间格式 to 帧 and the node divides your typed numbers by 30 to print seconds - so you enter frames, and the total-duration line gets the same unit stuck on it. Pick one unit per run; mixing seconds and frames across shots is how you end up with a shot list that runs backwards.

The default shot times overrun the length limit. The shipped defaults run 0–3, 4–6, 7–10, 11–13 and 14–16 seconds, while 视频总时长 caps at 15. With four or five shots you should retime them by hand so the last shot ends inside the total you declared - otherwise your prompt contains two contradictory durations, and the model gets to pick.

API calls cost per run. Wan 2.6 is API-only, and video is the expensive corner of the metered-node world. Build and test the prompt with the structural table in front of you before you send it.