📢 UPDATE: Typhoon V2 has landed!
Now with proper 768px-native training, higher-res LoRAs, and fewer anatomical crimes.
A personal labor of chaotic love — Typhoon is a finely merged and experimentally-trained Stable Diffusion 1.5 model, balancing character fidelity, facial detail, and a very specific aesthetic sensibility. It’s been trained, broken, retrained, merged, unmerged, cried over, and eventually released into the wild. Now you get to enjoy it.
Note: This is the SD1.5 version. Same soul, slightly different tempers across versions.
Want to see Typhoon in action with a fully customized WebUI?
Check out CipherCore-SD-1.5-WebUI — a clean, fast interface built around Typhoon as the default model:
👉 Raxephion/CipherCore-SD-1.5-WebUI
This repo includes:
- Built-in Typhoon support (all versions)
- Optimized SD1.5 workflows
- Clean UI with minimal overhead
- Out-of-the-box settings tailored to the model
- Optimised for CPU
- Free for personal offline, local use
Typhoon started out simple — until it wasn’t. After multiple failed attempts (3 out of 4 training runs, to be exact) and renting GPUs for way too long, what finally emerged was a model that holds up remarkably well in portrait and stylistic renders.
It was trained using a mix of:
- First-stage full checkpoint finetuning
- LoRA training specific to each aesthetic goal
- Merging those LoRAs back in with… patience, math, and caffeine
To help make sense of this beautiful mess, I also built two tools to analyze merge quality:
They're in beta. Which means math happens, but sometimes weirdly.
- Trained on 512x512 crops (regret lives here)
- Balanced, dreamy aesthetic with strong facial fidelity
- Merged with multiple LoRAs designed for portraiture and detail
- Performs surprisingly well across samplers and CFGs
- Still great for tag-heavy prompting
📸 Same seed. Same settings. Noticeably sharper.
Typhoon V2 is a full refresh:
- ✅ 768px-native training (finally!)
- ✅ Entire dataset reprocessed, upscaled, and tagged better
- ✅ Cleaner anatomy, facial symmetry, and light handling
- ✅ Merged with new, high-res LoRAs trained from scratch
- ✅ More reliable with consistency + stylized fidelity
- ✅ Still tag-friendly — same prompting style applies
It retains the soul of V1 — but with fewer oddities, stronger detail retention, and much better high-res handling. You can drop it into your current workflows and expect it to just… work better.
V2 is now the recommended default.
- Resolution: 512x768 (or 576x768 for a slightly broader feel)
- CFG Scale: 7 (experiment 0.3–0.8 for softer output)
- Sampler: DPM++ 2M (Karras), Euler, or Euler a
- Hires Fix:
- Denoising: 0.7
- Upscale: 2x (Latent)
- CFG: 7
No aDetailer or face fix required. Model handles facial detail well.
- Same samplers and CFG work well
- Hires fix denoising can drop to 0.5–0.6 for best edge retention
- Native 768px input allows for better aspect ratios — 768x1152 or similar
Both versions of Typhoon respond best to tag-like prompts over natural language. The datasets were heavily captioned in this style, so stick to clean, concise tags for best results.
Examples:
masterpiece, 1girl, solo, detailed eyes, soft lighting, outdoorsportrait, close-up, shallow depth of field, 4k, photorealistic
Sample images below use hires fix as described.
No LoRAs applied — what you see is what the model outputs by default.
 |  |  |  |  |
 |  |  |  |  |
- NSFW results are hit or miss — partially due to neutered base model. V2 fares better here, but no promises.
- Tag-style prompting works best. Prose prompts may cause drift.
- Anatomy isn't perfect. But it’s improving. The fairy visits more often now.
This model is provided under a modified CreativeML Open RAIL-M license:
- ✅ Personal, private use is allowed and encouraged.
- ❌ Do not merge this model with other checkpoints or LoRAs — you’ll break the aesthetic.
- ❌ Do not upload to public generation sites or bots.
See the LICENSE file for the boring legal bit.
- Base: Stable Diffusion 1.5 (
v1-5-pruned-emaonly) - VAE: stabilityai/sd-vae-ft-ema
- LoRA analyzers: Raxephion Tools
- Diffusers Repo V1: Raxephion/Typhoon-SD1.5-V1
- Diffusers Repo V1: Raxephion/Typhoon-SD1.5-V2