varg podcast video

Generate consistent-character podcast and dialogue videos with varg image generation + image-to-video with native audio.

Generation runs on varg: the varg-ai skill (REST at https://api.varg.ai/v2, sign-in and model details) or the varg MCP server (https://mcp.varg.ai/mcp: generate_image, generate_video or call_tool, then get_job). The REST calls below map one-to-one onto the MCP tools.

When to use

• Two-character conversation videos (podcast, interview, debate)
• Educational content with a "lie vs truth" format
• Any multi-scene video where character consistency matters across angles
• Short-form content (5 s per scene, punchy dialogue lines)

Prerequisites

• A VARG_API_KEY (the varg-ai skill finds or sets one up) or the varg MCP connection
• Credits on the varg balance (GET /v2/billing/balance or MCP check_balance)
• Reference images for each character (photos or AI-generated portraits), uploaded with POST /v2/files so you have URLs

The workflow (8 steps)

Every generation returns a job. Poll GET https://api.varg.ai/v2/jobs/{id} until status is completed; the file is at output.outputs[0].url.

Step 1: Write a short script

Each scene = 5 seconds. One short punchy line per scene. No long monologues.

Scene 1.1 — Host: "So... is multitasking actually a superpower?"
Scene 1.2 — Trickster: "Absolutely. Five tabs, three calls, zero problems."
Scene 1.3 — Trickster: "Focus is overrated. Real pros do everything at once."
Scene 1.4 — Host (truth cutaway, to camera): "Truth: every switch costs focus. One thing at a time gets more done."

Format: Scene X.Y — Character: "short line". Keep each line to what its shot can say at about 2.5 words a second (a 5 s shot holds about 12 words; under 10 for a lip-synced close-up). Size and check them with the video-script skill.

Step 2: Generate the studio set (empty)

Generate the background/set as a standalone image with no characters. It becomes the base reference for every later shot.

curl -s -X POST https://api.varg.ai/v2/image \
  -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano_banana_2_1",
    "prompt": "A modern podcast studio interior, wide establishing shot. Dark walnut wood paneled walls, two brown leather armchairs facing each other at an angle, a small dark wood table between them with two black microphones on boom arms. Warm amber recessed lighting. Bookshelf background. No people. Empty studio set.",
    "aspect_ratio": "16:9"
  }'

Step 3: Generate character clothing (if needed)

If a character needs clothing different from their reference photo, generate the outfit as a standalone image:

curl -s -X POST https://api.varg.ai/v2/image \
  -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano_banana_2_1",
    "prompt": "A dark elegant Victorian-style outfit on a mannequin. Black and dark red velvet tailcoat, high collar, dark crimson vest. Gothic aristocratic aesthetic. No person, just the clothing. Studio product photography.",
    "aspect_ratio": "1:1"
  }'

Step 4: Generate the wide shot (both characters)

Pass all references together in files: character 1 + character 2 + clothing + studio. With files, nano_banana_2_1 edits from the references.

CRITICAL: Do NOT describe the characters' appearance in the prompt. Reference them by position ("left character from ref 1", "right character from ref 2 wearing outfit from ref 3"). Describing appearance conflicts with the ref image and produces a different person.

curl -s -X POST https://api.varg.ai/v2/image \
  -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano_banana_2_1",
    "prompt": "Two characters sitting in leather armchairs facing each other in this podcast studio. Left character from ref 1. Right character from ref 2 wearing the outfit from ref 3. Small table with microphones between them. Warm lighting.",
    "files": [
      { "url": "HOST_REF_URL" },
      { "url": "TRICKSTER_REF_URL" },
      { "url": "TRICKSTER_CLOTHES_URL" },
      { "url": "STUDIO_URL" }
    ],
    "aspect_ratio": "16:9"
  }'

This wide shot becomes the base reference for all camera angles.

Step 5: Generate camera angles from the wide shot

Pass the wide shot as the first reference. The prompt must say:

"Exact same scene as the reference image — same room, same characters, same positions, same lighting, same set. ONLY change: [camera move]. Do NOT rearrange anything. Do NOT move characters. Do NOT change the set."

# Close-up on character A
curl -s -X POST https://api.varg.ai/v2/image \
  -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano_banana_2_1",
    "prompt": "Exact same scene as the reference image — same room, same characters, same positions, same lighting, same everything. ONLY change: move camera closer to the left character. Same angle, just tighter framing. Do NOT rearrange anything. Do NOT move characters. Do NOT change the set.",
    "files": [
      { "url": "WIDE_SHOT_URL" },
      { "url": "CHARACTER_A_REF" }
    ],
    "aspect_ratio": "16:9"
  }'

# Close-up on character B
# Same, but "move camera closer to the right character" + pass character B's ref

# Truth cutaway (lighting change)
# "Exact same scene — ONLY changes: shift lighting from warm to cold blue, the left character turns their head to look directly at camera. Do NOT rearrange anything."

What works:

• Close-up on character A (camera moves closer)
• Close-up on character B (camera moves closer)
• Lighting shift (warm → cold blue)

What does NOT work:

• Over-the-shoulder shots (OTS): the model rearranges characters and breaks positions
• Describing new appearance details in the prompt: produces a different character

Show the stills to the user and get approval before any video. Stills cost a few credits each; video clips cost ten times more or more.

Step 6: Image-to-video with native audio

CRITICAL: Put the dialogue line directly in the video prompt and ask for native audio. The model voices the line itself; no separate TTS.

curl -s -X POST https://api.varg.ai/v2/video \
  -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kling_v4",
    "prompt": "The man in the grey sweater leans forward toward the microphone and says: \"So... is multitasking actually a superpower?\" He gestures with open hands, eyes widen with curiosity, natural talking motion.",
    "files": [{ "url": "CLOSEUP_HOST_URL" }],
    "duration": 5,
    "aspect_ratio": "16:9",
    "audio": true
  }'

Prompt structure for video:

[Character description] [action/gesture] and says: "[exact dialogue line]" [more action]

• "audio": true is varg's unified native-audio switch. Without it, or on a model without native audio, the clip is silent. A flat "provider_options": {"varg": {"generate_audio": true}} is dropped by the API.
• One image in files = the first frame. Each close-up from step 5 is the start of its scene.
• Model choice: kling_v4 (default here, native audio, integer 3–30 s); seedance_2_5 for hero close-ups (best faces and lip movement, higher price); heygen_video_1 for cheap drafts (sound is built in, 5–15 s). minimax_h3_max has no native audio on varg's current routes, so it comes back silent.
• Price one clip with POST /v2/estimate (same body) or MCP estimate_cost before the batch, quote it in credits, and wait for the user's go.

Add action to every scene; without action the video is a static talking head:

• Leans forward/back, gestures, picks up an object, waves a hand, crosses arms
• Snaps fingers, points, draws in the air, taps the table
• Eye roll, head tilt, eyebrow raise, a smirk forming

Voice drift. Native audio picks the voice per clip, so a character can sound different from scene to scene. If that happens, generate each character's lines with eleven_v4 and one fixed voice (POST /v2/speech), then lip-sync each close-up with sync_v3 (a silent clip + the line) or minimax_h3_max_lipsync (the still + the line).

Step 7: Build an HTML storyboard

Create an HTML file with all videos embedded. Include:

• A "Play all" button that plays the scenes in sequence
• Each scene: video (320×180) + character tag + dialogue line + duration
• Color coding: warm scenes (red border), truth cutaways (blue border)

<video src="VIDEO_URL" controls playsinline></video>

// Play all in sequence
function playAll() {
  let i = 0;
  function next() {
    if (i >= videos.length) return;
    videos.forEach(v => v.pause());
    const v = videos[i];
    v.currentTime = 0;
    v.muted = false;
    v.play();
    v.scrollIntoView({ behavior: 'smooth', block: 'center' });
    v.onended = () => { i++; next(); };
  }
  next();
}

Step 8: Stitch the approved scenes into one MP4

Locally, with the clips downloaded in order:

printf "file '%s'\n" scene_1_1.mp4 scene_1_2.mp4 scene_1_3.mp4 > list.txt
ffmpeg -f concat -safe 0 -i list.txt -c:v libx264 -c:a aac -r 30 episode.mp4

Or on varg, with the ffmpeg tool (POST /v2/ffmpeg, MCP call_tool with tool ffmpeg):

{ "model": "ffmpeg",
  "command": "-i {{in_1}} -i {{in_2}} -i {{in_3}} -filter_complex \"[0:v][0:a][1:v][1:a][2:v][2:a]concat=n=3:v=1:a=1[v][a]\" -map \"[v]\" -map \"[a]\" -c:v libx264 -c:a aac {{out_1}}",
  "input_files": { "in_1": "SCENE_1_1_URL", "in_2": "SCENE_1_2_URL", "in_3": "SCENE_1_3_URL" },
  "output_files": { "out_1": "episode.mp4" } }

Captions, if wanted: transcribe the episode (POST /v2/transcription, word timestamps included) and burn them in, or rebuild the episode as a cloud render with <Captions> (see varg-ai).

Key rules (learned the hard way)

1. Never describe appearance in edit prompts. Pass ref images and say "character from ref 1". Describing appearance = a different person.
2. The wide shot is the base for all angles. Pass it as the first ref and say "exact same scene, ONLY change the camera".
3. No OTS shots. The model rearranges characters. Use close-ups instead.
4. Dialogue goes in the video prompt, and the video model voices it.
5. "audio": true on a model with native audio; otherwise the video is silent.
6. 5 seconds per scene: short lines, one idea per scene. Long monologues don't fit.
7. Add physical action: gestures, objects, movement. Static talking heads are boring.
8. Generate clothes separately if a character's ref photo has the wrong outfit, and pass them as an extra ref.
9. Truth cutaway = lighting change only: same positions, just warm → blue + the character looks at camera.
10. Keep prompts byte-identical when re-running: identical requests are free cache hits.

Cost estimate

Rough credits from the varg catalog (October 2026); quote the exact figure from POST /v2/estimate / estimate_cost.

Video is almost all of the cost. heygen_video_1 drafts cost a fraction of that; seedance_2_5 costs more than twice as much.

File structure

podcast/
  script.md                 # Full script (all questions)
  storyboard.html           # HTML review page with embedded videos
  ref_host.png              # Character reference
  ref_trickster.png         # Character reference
  studio_set.png            # Studio background
  trickster_clothes.png     # Clothing reference
  wide.png                  # Wide shot (base for angles)
  cu_host.png               # Close-up, host
  cu_trickster.png          # Close-up, trickster
  truth.png                 # Truth cutaway (blue light)
  scenes/                   # Generated clips, one per scene
  episode.mp4               # Stitched result

Related skills

varg-ai (sign-in, models, REST and cloud render), varg-consistent-characters (the same people across several locations, with captions and QA), varg-character-keynote (one speaker on a stage), video-skills (find a skill for another format, e.g. clipping a real podcast).