nodecode-dashscope-video

DashScope video generation: short clips on Alibaba Cloud's HappyHorse models, from text, an image or a reference image. A Nodecode rewrite of rriggs/hermes-plugin-dashscope-video by rriggs.

The original is a Hermes backend for its video_generate tool: Alibaba Cloud Model Studio's video models, spoken to over DashScope's async task API. The cell keeps that work as one tool, video_generate, and one verb you can call through eval. A clip takes two to three minutes and is billed to your Alibaba Cloud account, so one call is one considered prompt.

Get it

Tell the agent "add the dashscope-video cell", or run nodecode add nc://dashscope-video.

Set it up

A key comes from Alibaba Cloud Model Studio's console. Token-plan keys start sk-sp- and answer on the token-plan endpoint; pay-as-you-go keys start sk-ws- and answer on https://dashscope-intl.aliyuncs.com. Put either in the dashscope-video section of config.jsonc (or /settings):

{
  "dashscope-video": {
    "api_key": "sk-sp-…",              // or leave it out and export QWEN_API_KEY
    "api": "https://token-plan.ap-southeast-1.maas.aliyuncs.com",
    "model_family": "happyhorse-1.1",   // -t2v, -i2v, -r2v are appended
    "duration": 5                        // seconds, 3 to 10
  }
}

key_env names the environment variable the key is read from, QWEN_API_KEY unless it says otherwise, so a key exported in a shell needs no section at all. A key is required: without one the cell refuses at start and names the variable it looked for. /dashscope-video answers the endpoint, the key it read, the model each mode asks for, and the last clip.

Use it

The model calls video_generate; you can call the same work from eval:

video_generate with prompt — text-to-video. The model id is the family plus -t2v.
video_generate with prompt and image_url — image-to-video (-i2v). The image may be an https url, a data: uri, or a local path, which the cell inlines as a base64 data uri (up to 10 MB).
video_generate with prompt and reference_image_urls — reference-to-video (-r2v): the clip keeps the subject of the images you name. The family takes one, so the first is used.
duration is 3 to 10 seconds and aspect_ratio is 16:9, 9:16, 1:1, 4:3 or 3:4; out-of-range lengths are clamped, an unknown shape is 16:9.
model takes a full model id and overrides the family for that call.

(dashscope-video:generate "a lighthouse in fog, slow push in" :duration 5)
(dashscope-video:generate "she walks on" :image-url "~/frames/first.png"
                                          :aspect-ratio "9:16")

Both answer the same facts: the path of the saved clip, the url DashScope answered (it expires; the path does not), the model and mode used, the task id and the size. The clip is cached under this home's cache/videos/.

The model id for a mode resolves in the original's order: an explicit model argument, then the section's model_t2v / model_i2v / model_r2v, then model_family plus the mode's suffix.

Test it

NODECODE_SRC=~/nodecode/nodecode .github/test-cell.sh tools/nodecode-dashscope-video

Every request is a stubbed nlk:http and every wait a stubbed clock, so the suite reaches nothing and sleeps not at all.

Gaps

The original also accepted negative_prompt, audio, seed and resolution arguments on its provider; the family supports none of them and the request payload never carried them, so the cell's tool does not offer them. It has no model listing and no status command beyond /dashscope-video: the original had none either.