# THE BAKER — Flux.2 LoRA Training Summary

**Created:** 2026-05-13  
**Status:** ✅ Infrastructure ready, awaiting execution

---

## What Was Built

Complete fine-tuning pipeline for Flux.2 to generate THE BAKER deck visuals and concept art.

### Files Created (8 total)

| File | Size | Purpose |
|------|------|---------|
| `00_START_HERE.md` | 4 KB | Quick start guide |
| `00_quickstart.sh` | 5 KB | Interactive workflow |
| `README.md` | 6 KB | Full documentation |
| `TRAINING_PLAN.md` | 8 KB | Execution plan & QA |
| `config.json` | 2 KB | Training configuration |
| `generate_dataset.py` | 19 KB | Dataset generation (54 images) |
| `generate_deck_visual.py` | 9 KB | Deck visual generator (13 presets) |
| `train_flux_lora.py` | 14 KB | LoRA training script |
| `visual_concepts.md` | 8 KB | 8 visual categories documented |

**Total:** ~75 KB of scripts + documentation

---

## Training Dataset

**54 images across 8 categories:**

1. **Estate Interiors** (10) — Barakat estate, baroque, phoenician, decay
2. **Church and Ritual** (8) — Maronite Catholic, communion, reverence
3. **Nightclub and Violence** (8) — Opulence with basement, hidden danger
4. **Lebanon and Memory** (8) — Byblos, civil war, cemetery, exile
5. **Family and Food** (6) — Communion reception, gatherings, tension
6. **Character Portraits** (10) — Fredric, Magda, Billy, Vincent, Aida
7. **Symbolic Objects** (8) — Crucifix, gold Beretta, bread and blood
8. **Title Cards** (6) — Deck visuals, typography, dividers

**Captions:** All use trigger word "the baker film style" + style keywords

---

## Output Deliverables

### After Training:
- `outputs/lora_weights.safetensors` (~100-200 MB)
- Training checkpoints every 500 steps
- Training logs with loss metrics

### After Generation:
- 13 deck visuals (1024x576, 16:9 format)
- Ready for Typst deck integration
- Reusable for marketing materials

---

## Execution Commands

### Quick Start (Interactive):
```bash
cd /home/workspaces/thebaker/training
./00_quickstart.sh
# Choose option 4 for full pipeline
```

### Manual (Step-by-Step):
```bash
# Step 1: Generate dataset
python generate_dataset.py --config config.json

# Step 2: Train LoRA
python train_flux_lora.py --config config.json

# Step 3: Generate deck visuals
python generate_deck_visual.py --batch --lora outputs/lora_weights.safetensors
```

---

## Resource Requirements

**GPU:** NVIDIA RTX 6000 Ada (48GB) ✓ Available  
**VRAM Usage:** ~20-24GB during training  
**Disk Space:** ~25 GB (base model) + 500 MB (dataset/outputs)  
**Time:** 4-6 hours total
- Dataset generation: 30-60 min
- Training: 2-4 hours
- Deck visual generation: 15-30 min

---

## Visual Style Captured

**Core Aesthetic:**
- Grandeur with decay
- Chiaroscuro lighting
- Baroque Catholic weight
- Phoenician memory
- Lebanese diaspora
- Morally compromised luxury
- Cinematic composition

**Trigger Word:** `the baker film style`

---

## Next Steps

1. **Review** `00_START_HERE.md` for quick start
2. **Run** `./00_quickstart.sh` and choose option 4
3. **Monitor** training progress (2-4 hours)
4. **Review** generated deck visuals
5. **Integrate** with `assets/decks/THE_BAKER_Cannes_Deck.typ`
6. **Compile** deck to PDF for Cannes

---

## Success Criteria

- [ ] 54 training images generated with consistent quality
- [ ] Training loss decreases steadily to < 0.1
- [ ] LoRA weights saved successfully (< 200 MB)
- [ ] 13 deck visuals match "grandeur with decay" aesthetic
- [ ] Deck visuals integrate cleanly with Typst deck
- [ ] Final PDF ready for Cannes pitch

---

## Integration Path

### Option A: Direct Image Insertion
```typst
#image("../training/assets/decks/visuals/02_the_bargain.png", 
       width: 100%, fit: "cover")
```

### Option B: Visual Reference Cards
Create new slide type with generated images as mood references

### Option C: Background Textures
Generate subtle textures for slide backgrounds

---

## Documentation Links

- **Quick Start:** `00_START_HERE.md`
- **Full Guide:** `README.md`
- **Execution Plan:** `TRAINING_PLAN.md`
- **Visual Concepts:** `visual_concepts.md`
- **Configuration:** `config.json`

---

## Ready to Execute

All infrastructure is in place. The training pipeline is ready to run.

**Start command:**
```bash
cd /home/workspaces/thebaker/training
./00_quickstart.sh
```

**Recommendation:** Run option 4 (Full pipeline) to complete the entire workflow in one session.

---

**Note:** This is a one-time setup. Once trained, the LoRA can be reused indefinitely for:
- Deck updates
- Marketing materials
- Social media assets
- Press kits
- Investor presentations
- Cast/crew lookbooks