| # Examples — Indic Heritage Studio v2 |
|
|
| This directory contains pre-baked demo outputs for the project profile PDF |
| and the demo video. Run `python scripts/generate_demo_outputs.py` to |
| generate the full gallery on your dev box before recording the AMD demo. |
|
|
| ## Structure |
|
|
| ``` |
| examples/ |
| ├── README.md ← this file |
| ├── inputs/ ← source images used for style transfer / video / inpainting |
| │ ├── portrait.jpg |
| │ ├── landscape.jpg |
| │ └── portrait_mask.png |
| ├── 1_text_to_image/ ← 5 styles × 3 prompts × 3 seeds = 45 images |
| │ ├── madhubani/ |
| │ ├── warli/ |
| │ ├── pattachitra/ |
| │ ├── mughal/ |
| │ └── tanjore/ |
| ├── 2_style_transfer/ ← 5 styles × 2 input images = 10 styled images |
| │ ├── madhubani/ |
| │ ├── warli/ |
| │ └── … |
| ├── 3_image_to_video/ ← 5 styles × 25-frame MP4 = 5 videos |
| │ ├── madhubani.mp4 |
| │ └── … |
| ├── 5_inpainting/ ← 1 inpainting demo |
| └── 6_controlnet/ ← 1 ControlNet demo (canny + mughal) |
| ``` |
|
|
| ## Demo video script (3-5 min) |
|
|
| 1. **0:00 — Title card + intro (15s)** |
| - "Indic Heritage Studio — multimodal heritage art generation on AMD Radeon + ROCm" |
| 2. **0:15 — GPU verification (15s)** |
| - `rocm-smi` visible in terminal |
| - `python scripts/verify_rocm.py` shows ROCm ready |
| 3. **0:30 — Text → Image demo (60s)** |
| - Run: `python -m core.text_to_image --prompt "..." --style madhubani --size 1024` |
| - Show output PNG |
| - Mention: "SDXL 1.0 + per-style LoRA, 1024×1024 native" |
| 4. **1:30 — Style transfer demo (45s)** |
| - Run: `python -m core.style_transfer --image examples/inputs/portrait.jpg --style mughal` |
| - Show before/after |
| - Mention: "IP-Adapter XL on SDXL, 30 steps" |
| 5. **2:15 — Image → Video demo (45s)** |
| - Run: `python -m core.image_to_video --image outputs/demo_styled_mughal.png --out outputs/demo.mp4` |
| - Play the 25-frame MP4 |
| - Mention: "Stable Video Diffusion XT 1.1, 25 frames at 8 fps" |
| 6. **3:00 — Inpainting + ControlNet demo (45s)** |
| - Quick show of inpainting output |
| - Quick show of ControlNet (canny → mughal) |
| 7. **3:45 — Multi-GPU benchmark (30s)** |
| - `python scripts/benchmark.py --configs 1 2 4 8 --samples 4` |
| - Show scaling chart |
| 8. **4:15 — UI walkthrough (30s)** |
| - `python app.py` → walk through all 6 tabs in browser |
| 9. **4:45 — Closing (15s)** |
| - "Track 1, TeamIndicForge, Indic Heritage Studio — submitted." |
|
|
| ## Pre-baking strategy |
|
|
| Run `python scripts/generate_demo_outputs.py` on your 8×80GB NVIDIA dev box |
| in Week 2. This takes ~30 minutes and produces every demo asset. The AMD |
| demo recording (Aug 4) then only needs to *re-run* a few commands live for |
| the camera — the actual outputs already exist as fallback if AMD |
| generation fails. |
|
|
| ## Curating the gallery for the PDF |
|
|
| After generating, curate: |
| - Pick 1 best image per style for the PDF cover |
| - Pick 1 best video for the PDF embedded figure |
| - Pick the most visually striking ControlNet result |
| - Pick a 2×2 grid showing 4 styles for the architecture section |
|
|
