AI Podcast

AI Podcast Generator

Create multi-person talking head podcast videos using the inference.sh pipeline: portrait generation → TTS audio → avatar video → merge. Supports real humans (via Phota), 3D mascots, illustrated characters, and mixed casts.

Use when the user wants to create a podcast, talking head video, demo reel, promotional conversation, or any multi-speaker video content.

Pipeline Overview

Characters (images) → TTS (audio per turn) → Avatar (video per turn) → Merge (final video)

Process

Step 1: Character Creation

Choose the right tool per character type:

Training a Phota identity (optional but recommended for humans):

If you need a real human character with consistent identity across multiple angles and shots, train a Phota profile first:

belt app run phota/train --input '{
  "images": ["url1.jpg", "url2.jpg", ...],
  "wait": true
}' --save profile.json

• Requires 30-50 face images of the subject
• Training takes a few minutes with wait: true
• Returns a profile_id you then use in phota/generate as [[profile_id]] in prompts
• The profile is reusable forever — train once, generate unlimited shots

If you don't need cross-shot consistency (e.g. single-speaker video, one angle only), pruna/p-image is simpler and cheaper.

Character sheets first, podcast frames second:

1. Generate a character sheet (plain white background, multiple angles) for each character
2. Then place characters into the podcast studio setting using the sheet as reference

For branded characters (logo on clothing):

1. Generate the character with a plain version of the garment
2. Use phota/edit with the logo as a second reference image to add the logo
3. Always pass the logo image alongside character references when generating new angles

Step 2: Alternate Angles

Generate at least 2 angles per character for visual variety:

For close-ups, prompt for "tight framing, chest up, shallow depth of field" — not "turned to the side" (which just makes them look away).

Identity consistency rules:

• For real humans with a Phota profile: use phota/generate or phota/edit for new angles — Gemini does not preserve facial identity and will produce a different person
• For real humans without a Phota profile: try to generate all needed angles in one go with pruna/p-image, or consider training a Phota profile if you need many shots
• For mascots/illustrations: Gemini 3 Pro is fine, pass the established frame as reference

Framing rule: Use tight framing on individual speakers. Wide shots with multiple seats show empty chairs when only one person is on screen.

Step 3: QA Frames

Before proceeding, visually inspect all frames for:

• Extra people in the background
• Multiple microphones (should be single mic per shot)
• Wrong or distorted logos
• Inconsistent character identity across angles
• Weird artifacts (extra limbs, merged objects)

Fix issues before generating video — re-rendering video is the most expensive step in the pipeline.

Step 4: Write the Script

Rules for natural conversation:

• Write it like a real conversation, NOT like people reading ad copy in turns
• Include reactions ("wait, hold on", "that is wild"), interruptions, and follow-up questions
• Vary turn length — short reactions (1 sentence) mixed with longer explanations (2-3 sentences)
• The host should ask real questions, not set up obvious talking points
• Keep total duration target in mind: ~2.5 words/second for natural speech at 1.05x rate

Duration guide:

Step 5: Generate TTS Audio

Use inworld/text-to-speech-2 for each turn.

belt app run inworld/text-to-speech-2 --input '{
  "text": "...",
  "voice_id": "...",
  "speaking_rate": 1.05,
  "audio_encoding": "MP3"
}' --save output.json

Voice selection:

• Generate samples with the same line across candidate voices BEFORE committing
• Let the user listen and approve voices
• Good podcast voices: Tyler, Nate, Lauren, Kelsey, Naomi, Anjali (EN_US)
• Use inworld/text-to-speech-2:voices to list all available voices

Speaking rate:

• Default to 1.05 for natural podcast pacing
• Use 1.1 for short snappy reactions
• NEVER go below 1.0 — sounds slow and disengaging
• Keep rate consistent per character across all their turns

All TTS turns can run in parallel (cheap, fast ~2-8s each).

Step 6: Generate Video Clips

Use pruna/p-video-avatar for each turn.

belt app run pruna/p-video-avatar --input '{
  "image": "<character_frame_url>",
  "audio": "<tts_audio_url>",
  "resolution": "720p",
  "video_prompt": "..."
}' --save output.json

Critical: Run clips SEQUENTIALLY, not in parallel. Parallel runs hit the same GPU and cause CUDA OOM failures. Each clip takes 15-90s depending on audio length.

Angle assignment plan: Alternate between front and close-up shots across turns for visual variety. Example for 6 turns:

T1: Speaker A — front
T2: Speaker B — front
T3: Speaker C — front (or close-up)
T4: Speaker A — close-up
T5: Speaker B — close-up
T6: Speaker A — front

Step 7: Merge

Use infsh/media-merger to stitch all clips into the final video.

# Build input JSON
{
  "media_files": [
    {"file": "<clip1_url>"},
    {"file": "<clip2_url>"},
    ...
  ],
  "fps": 24,
  "output_format": "mp4"
}

belt app run infsh/media-merger --input merger_input.json --save final.json

Merger is free and takes 2-6 minutes depending on total duration.

Rules

1. Gemini does not preserve human facial identity — generating alternate angles of a real human with Gemini will produce a different person. For identity-consistent human shots, use Phota with a trained profile_id, or generate all angles in a single batch. This was learned after Gemini produced an entirely different face for a close-up that was supposed to match the front shot.

2. NEVER run p-video-avatar clips in parallel — they compete for GPU memory and fail with CUDA OOM. Run them sequentially. This was learned after 2 of 3 parallel runs failed.

3. NEVER set speaking_rate below 1.0 — it sounds artificial and disengaging. Default to 1.05. Learned from user feedback that 0.9 rate "felt weird and disengaging."

4. ALWAYS QA frames before generating video — video generation is the most expensive step in the pipeline. Catching a double mic or wrong logo in the image stage is cheap to fix. Catching it after video generation means re-rendering the entire clip.

5. ALWAYS use tight framing for individual speaker shots — wide/establishing shots show empty seats where other speakers should be. Frame from waist or chest up so no empty chairs are visible.

6. ALWAYS pass the logo as a reference image when generating branded characters — describing a logo in text produces wrong results. Pass the actual logo file as a second image input.

7. ALWAYS get voice approval before full production — generate samples with the same line across 5-8 candidate voices and let the user pick before committing to the full script.

8. Script should read like a conversation, not an ad — people reading ad copy in turns sounds fake. Include reactions, interruptions, varied turn lengths, and genuine questions. The host should have personality, not just set up talking points.

App Reference

Use belt task cost <task-id> to check the cost of any individual task.