# 🌪️ Typhoon V1 (Stable Diffusion XL Edition)

> _"What began as Tornado V3 spun out into its own category-five situation."_

---

## 🧬 Overview

Typhoon SDXL is a handcrafted, stylized realism model trained from scratch on SDXL 1.0. It’s not a merge. It’s not raw realism. It’s something in-between — vivid, cinematic, and unapologetically polished.

What sets it apart is its lighting depth, sharp portraits, and ability to handle both natural language *and* tag-style prompts. That, and the fact that it required multiple GPU meltdowns and a stack of failed runs to get here.

Think of it as clean realism with opinions. It won’t just show you what’s there — it’ll make it look good doing it.

---

## 🔧 Development Notes

This wasn't a quick merge job or a comfy pipeline export. Typhoon SDXL was trained from a clean SDXL 1.0 base with:

- Full SDXL training (no baked junk)
- Curated subject+lighting datasets
- Custom LoRAs trained for style, anatomy, lighting, and facial sharpness
- Manual merge tuning with different weights per LoRA — lots of trial, lots of error

Along the way, I wrote tools to survive the chaos:

- [LoRA Strength Analyser](https://github.com/Raxephion/loRA-Strength-Analyser)  
- [LoRA Epoch Analyser](https://github.com/Raxephion/loRA-Epoch-Analyser)

Both are now being integrated into a Gradio UI, but for now — command line or bust.

> 1 in 4 training runs failed. The storm was literal.

---

## 🖼️ Sample Images

All examples are straight from Typhoon SDXL — no external LoRAs, no Adetailer, no fixers.

**Settings:**

- **Resolution**: 1024×1024  
- **Sampler**: DPM++ 2M Karras (Euler, Euler A, and SDE also fine)  
- **Steps**: 15–30  
- **CFG**: 0.5–0.7 (0.3–0.8 range works)  
- **Face Restoration**: Off. It ruins detail.  
- **VAE**: [stabilityai/sdxl-vae](https://huggingface.co/stabilityai/sdxl-vae)

---

## ⚙️ Prompting Tips

- **Trigger Words**: Not needed  
- **Prompt Style**: Both natural language and tag-style work  
- **For Best Results**:  
  - Use short, vivid descriptions for stylized realism  
  - Use tag-style for concept art / illustration  
- **Close-ups**: Excellent by default — no Adetailer required  
- **Scene Balance**: Too much gloss or lighting can oversaturate skin/fabric — go easy on lighting boosts

---

## ✨ Style Guide

Typhoon SDXL excels at:

- Portraits and close-ups  
- Stylized realism / cinematic concept art  
- Balanced lighting, clean renders  
- Semi-fantasy aesthetics or “rendered real” styles

It **can** do gritty realism — but it won’t default there. You’ll need to prompt it hard.

---

## ⚠️ Limitations

- **NSFW**: Heavily mixed results — restoration was partial  
- **Realism**: Defaults to stylized, needs prompting to go raw  
- **High-Gloss Effects**: Can overdo lighting on skin/fabric  
- **Prompt Drift**: Like all SDXL models, badly structured prompts may wander

---

## 🔒 License & Usage

- ✅ Personal/local use encouraged  
- 🚫 **Do NOT** upload to generation bots/sites  
- 🚫 **Do NOT** merge this model with others  

> Typhoon’s internal balance is LoRA-weight tuned — merges will break the style and waste the work. Keep it local. Keep it clean.

---

## 🔮 Future Work

Typhoon V2 will include:

- Larger dataset coverage  
- Improved anatomy + realism control  
- More stable NSFW behavior (if possible)  
- Further tuning tools and patch updates

I’m also integrating my analysis tools into a web UI for easier merging, tuning, and model dissection.

---

🌩️ ☕ Support the Storm

If you like what I’m building — Typhoon, Tornado, the tools, the chaos — and want to help keep it all spinning, consider supporting me on Ko-fi:
https://ko-fi.com/raxephion

Every bit helps cover compute costs, caffeine, and the occasional "why is this broken?" meltdown. Thanks for keeping the storm alive.

---

**Enjoy the storm.** ⛈️