Wan25图生视频

Fourteen input fields sounds like a lot for a prompt. It is, and that's the point: image-to-video prompting on Wan is prose plus separated motion, camera and audio directives, and typing all of that by hand for every take gets old by the third clip.

What it is, and where it fits

Wan25图生视频 takes your scene, motion, camera and sound choices and emits three ready-to-use strings: a visual prompt, a sound prompt, and a complete prompt with both folded together. It calls nothing and needs no key - it's a text node. The author's own note is short and worth heeding: use the complete prompt.

Important context, because it shapes who this is for. Wan 2.5 has no open weights - 2.1 and 2.2 are Apache 2.0 and still the base everything builds on, but from 2.5 onward Alibaba put the numbered series behind an API. So this node targets the Wan 2.5 API node, not a local GGUF, and 2.5 is where the series picked up native audio, which is why half these fields are about sound.

The useful consequence: the visual half is normal prose prompting, which is how Wan has always wanted to be prompted. You can point 视觉提示词 at a local Wan 2.2 workflow and get good results. The sound fields will just sit there doing nothing, because no version of open Wan generates audio at all.

How it builds the three outputs

The visual string is a comma-joined assembly line: your scene description, then the subject action, then the environment-effect description, then the camera sentence, then a visual-consistency clause, then your extra notes. The camera sentence is where the structure shows - the movement's description has the target substituted into it, then gets a speed word and a duration:

画面描述，主体自然站立，环境动效描述，镜头推近主体，速度适中，历时3秒，保持风格统一

The sound string is the same idea with fewer parts: your vocal/voice description, plus either 音效：<your detail> or a generic 添加环境音效, plus the BGM style if you picked one.

The complete prompt is just the two of them labelled - 【视觉】… 【声音】…. That bracketed two-block shape is the point: on LLM-encoded models, clean separation beats clever syntax.

The inputs that matter

All fourteen are required, so there's no blank-slate mode - but only four or five change the output meaningfully:

• 画面描述 - the scene. The default is placeholder text; replace it first.
• 主体动作 - a verb phrase. The node prefixes it with 主体 when assembling, so 自然站立 becomes 主体自然站立. Write 转身走向门口, not 女孩转身走向门口, or you'll get 主体女孩转身走向门口.
• 运镜方式 + 镜头目标 - 35 movements, and the target string gets substituted into the movement description. The default 主体 is generic; if your shot is about a specific person or object, name it here.
• 声音描述 - the vocal half. Default is 清晰的对话声; clear it if the clip is silent-ish.
• 是否添加音效 - 是/否. On 是 with an empty 音效细节 you get the generic 环境音效 line; fill 音效细节 for anything specific.

Then the knobs: 场景动效 (44 options, default 无, with 环境动效 marked 推荐 - a good cheap default, it gives the frame life without inventing action), 运镜速度 and 运镜时长, 视觉连贯性 (高度一致 / 风格统一 / 色彩渐变 / 自由创作), 背景音乐, and a free-text 附加描述. Duration sliders across the pack now run 1–15s in one-second steps.

Install

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

Or Manager → search prompt-formula. Restart, then find it under 📕提示词公式 / 图转视频. Zero dependencies beyond the torch>=2.0.0 line you already satisfy, and zero model downloads.

Gotchas

There's no image input. Every one of the fourteen fields is text. The image goes into the Wan 2.5 API node (or your local Wan I2V sampler); this node only writes the prompt. If you expected to feed it a picture, that assumption is the bug.

The audio fields only do something on the 2.5 API path. On a local Wan 2.2 workflow, output 2 and the whole sound half of output 3 are going straight into a model that has no audio branch. Don't spend ten minutes tuning 背景音乐 for a local render - reach for a Foley stage or an audio-video model instead.

Local Wan is slow, and prompting won't fix that. A five-second clip is commonly 15–40 minutes on a 4090 without speed LoRAs, plus a 12–16GB VRAM floor for the 14B family. Better prompts improve the odds of a usable take, not throughput.

Everything blank means nothing out. Empty fields and 无 selections simply drop out of the assembled string, so a fully-cleared node returns an empty prompt rather than an error and your sampler gets a blank. If a render comes back as a random-looking mess, read the string before you blame the seed.

API calls are metered. Video is where per-call cost escalates fastest, and the standing community grumble about API nodes is exactly that they turn a free-after-electricity render into a bill. Test at low settings before you spend on a full run.