docs/DEMO_SCRIPT.md

Demo Recording Script

The final submission should use the AMD GPU version if credits arrive in time. Record the draft now so the story, screen flow, and timing are ready.

Draft Demo Before AMD Credits

Target length: 2-3 minutes.

Scene 1 - Problem

Show the title slide or Hugging Face Space.

Narration:

"Long-form creators need short clips for TikTok, Shorts, and Reels, but finding the best moments, trimming them, and adding subtitles can take hours."

Scene 2 - Channel Profile

Open the local web app at http://localhost:5173.

Set:

• Niche: education, gaming, or podcast
• Clip style: funny or informative
• Clip length: 60
• Language: Thai
• Platform: tiktok

Narration:

"ElevenClip.AI starts with a reusable channel profile, so highlight selection matches the creator's niche, language, style, and target platform."

Scene 3 - Video Input

Upload a short MP4 file or paste a YouTube URL.

Narration:

"The app accepts uploaded files or YouTube URLs. In production, the backend downloads the source, transcribes it, ranks highlights, and renders clips."

Scene 4 - AI Pipeline

Click Start Pipeline and show progress.

Narration:

"The AMD version runs Whisper Large V3 for transcription, Qwen2.5 for profile-aware highlight scoring, optional Qwen2-VL for visual signals, and ffmpeg for clip generation."

Scene 5 - Human-AI Editor

Show generated clips.

Edit a subtitle line, change start/end, approve a clip, and click download.

Narration:

"AI creates the first pass, but the creator stays in control. They can trim clips, edit subtitles, delete weak clips, regenerate a specific clip, approve, and download."

Scene 6 - AMD GPU Plan

Show README or benchmark placeholder.

Narration:

"Once AMD Developer Cloud credits are active, the same pipeline runs on ROCm and AMD Instinct MI300X. We will benchmark CPU versus AMD GPU and target ten subtitled clips from a two-hour video in under ten minutes."

Final Demo After AMD Credits

Target length: 3-4 minutes.

Add these shots to the draft demo:

1. AMD Developer Cloud instance page.
2. Terminal showing ROCm/GPU visibility:

rocminfo | head
python - <<'PY'
import torch
print(torch.cuda.is_available())
print(torch.cuda.get_device_name(0))
print(torch.version.hip)
PY

3. Backend running with DEMO_MODE=false.
4. Benchmark command:

python scripts/benchmark.py --youtube-url "<demo-video-url>" --language Thai --style informative --niche education --clip-length 60

5. Timing JSON showing input, transcription, highlight_detection, clip_generation, and total.
6. Final clips shown in the editor.

Recording Checklist

• Browser zoom at 100%.
• Use a short clean video for draft demo.
• Hide private tokens and email inboxes.
• Keep terminal text large enough to read.
• Record in 1080p or higher.
• End with GitHub URL, Hugging Face Space URL, and AMD/ROCm benchmark result.