🌪️ Typhoon V1 (Stable Diffusion 1.5 Edition)
"What began as Tornado V3 spun out into its own category-five situation."
🧬 Overview
Typhoon V1 for SD1.5 is a custom-trained model with a strong stylistic focus, built from scratch using carefully curated and tagged datasets. Initially meant as a continuation of Tornado, it soon veered off into its own trajectory — aesthetically and technically — and deserved a new name (and its own emotional baggage).
Despite the limitations of the SD1.5 architecture, Typhoon V1 inherits the stylization strengths of its SDXL sibling: sharp close-ups, expressive faces, confident eye rendering, and strong composition. It doesn't need ADetailer for faces and generally does a solid job on its own — unless something weird sneaks in.
Just don't expect poetic prose prompts to work miracles — this model responds best to short, tag-like prompts. That's because the dataset was trained using tagged images rather than natural language descriptions.
🔧 Development Notes
Typhoon was trained using a combination of full checkpoint training and LoRA modules, merged carefully one by one. Without local hardware capable of full-scale training, GPU rentals were required — and with 3 out of 4 training runs failing, this project was not particularly budget-friendly.
Merging the LoRAs involved trial-and-error with strength tuning, weight adjustments, and the occasional expletive. To make things slightly less chaotic, I wrote a Python tool called the LoRA Strength & Epoch Analyser, available below:
Choosing a base model also proved tricky. Many SD1.5 bases are overly refined, biased, or neutered. I eventually settled on v1-5-pruned-emaonly.safetensors, which was the least problematic. That said, its limitations still show through.
Oh — and yes, all training datasets were cropped to 512×512. This turned out to be a mistake, as it likely contributed to some of the model's anatomy and artifact issues. I’ll be correcting that for V2 when I scale the datasets properly. Eventually.
🖼️ Sample Images
All images were generated without LoRAs, using the base Typhoon V1 model. Post-processing includes Hires Fix (if that qualifies as post-processing).
Settings:
- Resolution: 512×768 or 576×768
- Sampler: DPM++ 2M Karras (Euler and Euler A also work)
- CFG: 7
- Hires Fix:
- Denoising strength: 0.7
- Upscaler: Latent
- Upscale by: 2
- Hires CFG: 7
- VAE: sd-vae-ft-ema
⚠️ Legacy .vae.bin or .vae.pt files may cause washed-out or desaturated results. The official VAE or none at all is strongly recommended — they yield identical (correct) results.
⚙️ Prompting Tips
- Trigger Words: None required
- Prompting Style: Short, tag-style prompts preferred (e.g.
1girl, blue eyes, looking at viewer)
- Natural Language: Avoid it — this model thinks in tags, not sentences
- ADetailer / Face Fixing: Not needed unless there's trouble with anatomy
- Recommended Resolutions:
- 512×768
- 576×768
- 640×832
Avoid narrower resolutions like 512×640 unless you enjoy triple elbows and bonus fingers.
⚠️ Limitations
- NSFW: Base model is partially neutered. Typhoon tries its best, but results vary. When it clicks, it really works — but don't expect consistency in that domain.
- Anatomy & Artifacts: Can struggle at narrow aspect ratios. Use Hires Fix whenever possible.
- Natural Language Prompts: Doesn’t like them. Stick to short, structured prompts for best results.
🔒 License & Usage
- ✅ Personal use: Absolutely
- 🚫 Do NOT upload this model to generation sites (e.g. online prompt tools or aggregators)
- 🚫 Do NOT merge this model into other models
Why? Typhoon’s structure is the result of a very delicate merge process involving multiple LoRAs, adjusted manually. Combining it with other models will throw off its aesthetic balance and likely degrade quality. Plus, I spent way too much on GPU rentals and debugging for it to become someone's merge ingredient. Please respect the effort.
🔮 Future Work
Typhoon V2 is planned, and will address:
- Properly scaled datasets (not just 512×512 crops)
- Improved base model (if a truly vanilla one can be sourced)
- Enhanced anatomy, reduced artifacts, and overall polish
Ongoing patch updates may be issued for V1 as more flaws are uncovered or edge cases appear.
🌩️ ☕ 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/raxephionEvery bit helps cover compute costs, caffeine, and the occasional "why is this broken?" meltdown. Thanks for keeping the storm alive.
Enjoy the storm. ⛈️
