# Training / Fine-tuning Roadmap

## Phase 1: Product first
Use ComfyUI + Wan/Hunyuan/CogVideoX to generate videos without training.

## Phase 2: Style LoRA
Collect 500-5,000 short clips in one niche, caption every clip, then train LoRA/style adapters.
Good niches:
- affiliate marketing shorts
- ecommerce product videos
- healthcare explainer shorts
- real estate walkthrough clips

## Phase 3: Domain model
After revenue, train deeper adapters or full fine-tunes using rented GPU clusters.

## Data rules
Only use videos you own, licensed stock videos, synthetic clips you generated, or Creative Commons content that allows this use.

## Hardware expectation
- Testing: RTX 4090 / A5000 / A6000 depending on model
- Serious generation: A100/H100 class GPU
- Full training from scratch: very expensive; avoid at start