🌪️ Typhoon V2 (Stable Diffusion 1.5 Edition)

"Still SD1.5. Still cursed. But at least now it understands limbs."

🧬 Overview

Typhoon V2 is the long-overdue upgrade to Typhoon V1, trained for SD1.5 using smarter techniques, cleaner data, and a few hard-earned lessons from the first version. The result? More control, better anatomy, stronger stylization — and fewer existential crises per batch.

It builds on the core identity of V1 (tag-based prompts, no trigger words, bold composition) but addresses its flaws head-on: warped limbs, prompt misfires, and the occasional brush with the uncanny.

V2 plays much nicer with short prompts, handles faces even better than before, and generally won’t lose its mind when asked for basic body parts. Still no poetic-prose magic, though — this one also thinks in tags.

🔧 Development Notes

Typhoon V2 was trained from scratch again, but this time with properly scaled datasets (no more 512×512-only crops) and aspect-ratio-aware augmentation. The architecture is still SD1.5, but the dataset strategy and training parameters got a much-needed overhaul.

Training was done on rented A100s — because, apparently, learning costs pain. Dataset prep was completely redone, with better captions, refined tag filtering, and entirely new sets focused on pose coverage, negative regularization, and structural accuracy.

Merging? None this time — this is a native checkpoint. No LoRA layering, no weight juggling. Just clean, consolidated training. (Still used my analysis tools, though — because mistakes are expensive.)

🛠️ Tools used during development:

LoRA Strength Analyser
LoRA Epoch Analyser
TensorPeek — for inspecting .safetensors metadata
LoRA Distiller (WIP) — experimental, but useful in this pipeline

The base model was once again v1-5-pruned-emaonly.safetensors, but augmented via LoRA distillation and pre-conditioning to mitigate its quirks. All improvements are checkpoint-native — no merging required.

🖼️ Sample Images

All images were generated using the base Typhoon V2 checkpoint. No LoRAs, no inpainting, no face fixers — just raw inference with Hires Fix.

Settings:

Resolution: 512×768, 576×832, or 640×896
Sampler: DPM++ 2M Karras (Euler A also works fine)
CFG: 6.5–7
Hires Fix:
- Denoising strength: 0.6–0.7
- Upscaler: Latent
- Upscale by: 2
VAE: sd-vae-ft-ema

⚠️ Legacy .vae.pt or .vae.bin files will likely cause washed-out or low-contrast results. Use the official VAE or none at all for correct output.

⚙️ Prompting Tips

Trigger Words: None
Prompting Style: Tag-based preferred (e.g. 1girl, long hair, looking at viewer)
Natural Language: Still not a fan — use structured tags for best results
ADetailer / Face Fixing: Rarely needed; faces are stable out of the box
Recommended Resolutions:
- 512×768
- 576×832
- 640×896
  Narrow resolutions like 512×640 are prone to artifacts. Avoid if possible.

⚠️ Limitations

NSFW: Still mildly neutered by the base model. Performance has improved over V1, but results are hit-or-miss.
Anatomy: Much better than V1, but still SD1.5 — expect occasional hiccups
Natural Language Prompts: Works better than V1, but short, tag-like prompts still yield the most consistent results

🔒 License & Usage

✅ Personal use: Absolutely
🚫 Do NOT upload this model to generation websites or aggregators
🚫 Do NOT merge this model into other checkpoints

Why? Typhoon V2 was trained cleanly and directly. Merging would break its stylistic balance, ruin its improvements, and waste the training effort. Please don’t.

🔮 Future Work

Further refinements possible via targeted LoRAs or partial retrains
Potential distillation into an SDXL variant
Ongoing prompt testing and edge-case analysis

🌩️ ☕ 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 — again. ⛈️

Raxephion
/

Typhoon-SD15-V2