varg-character-keynote

varg character keynote

One varg render produces a 9:16 talking-speaker video: Soul generates the character, Nano Banana 2.1 the empty stage, and Seedance 2.5 puts the character on that stage and voices the line with native audio.

Generation runs on varg: the varg MCP server (https://mcp.varg.ai/mcp, tool render_video) or the varg-ai skill (cloud render, POST /v2/render). Read varg-ai for sign-in, models and the render format.

To create or swap a character, edit ONLY three things

1. LINE: one or two short spoken sentences, emotional hook + payoff (fits 8 s).
2. character prompt: "Candid photo of ..." + body/face/clothing details + "Vivid saturated colors, realistic skin detail." Contrast sells the line: an exact age, an unusual build or a loud outfit played straight beats a generic presenter. The varg-character-design skill has archetypes, realism wording and moderation-safe phrasing; if the speaker must reappear in other videos, make a sheet there and pass its face crop instead of generating with Soul.
3. scene prompt: stage description, always "empty, no people", with a themed screen behind the speaker.

Template (cloud render: no imports)

const LINE = "I learned to code at sixty-eight. Your excuse just expired.";

const character = Image({
  prompt: "Candid photo of <PERSON: age, build, face, hair, clothing, accessories>. Vivid saturated colors, realistic skin detail.",
  model: varg.imageModel("soul"),
  aspectRatio: "9:16",
  // Soul "Realistic" style. Provider options must sit under varg.<provider>, or they are dropped.
  providerOptions: { varg: { higgsfield: { params: { style_id: "1cb4b936-77bf-4f9a-9039-f3d349a4cdbe" } } } },
});

const scene = Image({
  prompt: "Wide shot of a bright modern conference stage, empty, no people: dark backdrop with a large glowing screen showing <THEMED MINIMALIST SLIDE>, warm spotlights, standing microphone, blurred audience in the dark foreground. Vivid saturated colors, photorealistic event photography.",
  model: varg.imageModel("nano_banana_2_1"),
  aspectRatio: "9:16",
});

const talk = Video({
  prompt: {
    text: "The <person> from the first reference image stands center stage as a keynote speaker on the conference stage from the second reference image, holding the microphone. <One line of speaking manner/emotion.> Dialogue: \"" + LINE + "\" On the last words <one physical punchline action>, the audience erupts in applause. Warm spotlights, big-conference keynote energy, vivid saturated colors, no text on screen.",
    images: [character, scene],
  },
  model: varg.videoModel("seedance_2_5"),
  aspectRatio: "9:16",
  duration: 8,
  audio: "native",
});

export default (
  <Render width={1080} height={1920}>
    <Clip duration={8}>{talk}</Clip>
  </Render>
);

Run it

• MCP: price it with estimate_cost (tool render, input { "code": <template> }), then pass the template as code to render_video and poll get_job until it completes. Report the billed cost with get_job_cost.
• API: save it as keynote.tsx, then

curl -s -X POST https://api.varg.ai/v2/render \
  -H "Authorization: Bearer $VARG_API_KEY" -H "Content-Type: application/json" \
  -d "{\"code\": $(jq -Rs . < keynote.tsx)}"
# -> {"id":"job_xxx",...}; poll until completed, the video is at output.outputs[0].url
curl -s https://api.varg.ai/v2/jobs/JOB_ID -H "Authorization: Bearer $VARG_API_KEY"

Rules

• Cloud render code has NO import statements: Render, Clip, Image, Video and varg are globals. Models go ONLY through varg.imageModel() / varg.videoModel(); bare model strings crash.
• Image() and Video() are function calls, composed with JSX (<Clip>{talk}</Clip>).
• One clip = one continuous action; the physical punchline lands on the last words of LINE.
• Keep "no text on screen" and audio: "native": speech and applause come from Seedance's native audio.
• Price first. Run estimate_cost (MCP, tool render) or POST /v2/estimate on the Seedance request and quote the result in credits. One run is roughly 350 credits at Seedance's default 720p (catalog, October 2026) and takes about 5–15 minutes. Never launch a render without the user's explicit go-ahead.
• Seedance renders 720p by default. For 1080p, set the resolution for the route that runs it (see https://varg.ai/models/seedance_2_5.md); it roughly doubles the cost.
• aspectRatio must match on both Images, the Video and the Render canvas (1080×1920 = 9:16). Change all four together when changing format.
• Unchanged prompts are cached for free: when iterating on the line or the punchline, keep the character and scene prompts byte-identical so only the video regenerates.

Related skills

varg-ai (sign-in, models, cloud render format), varg-consistent-characters (the same people across several scenes), varg-podcast-video (two characters talking), video-skills (find a skill for another format).