# Plan: Train Model Tạo Ảnh 4K

## Mục tiêu
Tạo model tạo ảnh chất lượng cao native 2K/4K, self-host trên VPS H100, bán API.

## Infra
- VPS: Azure, 80 vCPU, 629GB RAM, 2x NVIDIA H100 NVL 96GB
- Số lượng VPS: 3+ (có thể scale thêm)
- GPU Driver: 580.95.05, CUDA 13.0

## Architecture: Cascaded Pipeline

```
Stage 1: Flux (fine-tuned) → Generate 1024px
Stage 2: SUPIR/StableSR (fine-tuned) → Upscale 1K → 2K
Stage 3: SUPIR/StableSR (fine-tuned) → Upscale 2K → 4K
```

## Cost ước tính
- Data collection: $0 (crawl free) + ~$20K (Opus caption, tuỳ deal)
- Training: $0 (VPS free)
- Tổng: $0 - $20K

## Timeline: ~8-9 tuần

| Phase | Công việc | Thời gian |
|-------|-----------|-----------|
| Phase 1 | Setup môi trường | 2-3 ngày |
| Phase 2 | Data collection | 1-2 tuần |
| Phase 3 | Training (3 stages song song) | 4-6 tuần |
| Phase 4 | Evaluation & iteration | 1-2 tuần |
| Phase 5 | Serving & API | 3-5 ngày |

---

## Phase 1: Setup môi trường

### Mỗi VPS cài:
- Python 3.10+
- PyTorch 2.x (CUDA 13.0)
- diffusers, transformers, accelerate, deepspeed
- img2dataset, webdataset
- SUPIR / StableSR dependencies

### Phân VPS:
- VPS-1: Data collection + Train Stage 1 (Flux)
- VPS-2: Train Stage 2 (SR 1K→2K)
- VPS-3: Train Stage 3 (SR 2K→4K)

---

## Phase 2: Data Collection

### 2A — Dataset Stage 1 (1-2M ảnh 1024px + caption)
- Crawl: Unsplash, Pexels, LAION-Aesthetics (score > 6.0)
- Caption: Opus 4.6 API
- Format: WebDataset (tar shards)

### 2B — Dataset Stage 2-3 (200K-500K ảnh 4K pairs)
- Crawl: Unsplash 4K, Flickr CC high-res
- Tạo pairs: downscale 4K → 2K → 1K
- Augmentation: crop, flip, color jitter, degradation

---

## Phase 3: Training

### Stage 1: Fine-tune Flux
- Base: Flux Dev / Schnell
- Method: LoRA rank 128 hoặc full fine-tune (DeepSpeed ZeRO-2)
- Batch size: 8-16 (2x H100)
- LR: 1e-5, cosine decay
- Steps: 100K-500K
- Resolution: 1024x1024

### Stage 2: SR 1K→2K
- Base: SUPIR / StableSR
- Loss: L1 + perceptual + GAN
- Batch size: 2-4 per GPU
- Steps: 200K-500K

### Stage 3: SR 2K→4K
- Base: SUPIR / StableSR
- Loss: L1 + perceptual + GAN
- Batch size: 1-2 per GPU
- Steps: 200K-500K

---

## Phase 4: Evaluation
- FID score, CLIP score
- Human eval
- So sánh với Midjourney, DALL-E 3, Flux Pro
- Iterate nếu chưa đạt

---

## Phase 5: Serving
- FastAPI + Triton Inference Server
- Queue: Redis/RabbitMQ
- Pipeline: Stage 1 → 2 → 3
- API: POST /generate, GET /status, GET /result
- TensorRT optimize, target < 30s/ảnh 4K