📢 Welcome to Typhoon SDXL — a stylized, cinematic model built for clarity, precision, and the occasional artful overkill.
This is not a raw realism model. It’s got opinions. Strong ones.
Typhoon SDXL began as a planned sequel to Tornado — but things escalated. What started as Tornado v3 took on its own identity, eventually diverging into a standalone model with a distinct, stylized realism that balances cinematic lighting, visual depth, and highly clean detail.
This model was built from the ground up using:
- ✅ Full SDXL 1.0 training (not a merge)
- ✅ Clean, curated datasets with subject + lighting emphasis
- ✅ A suite of custom-trained LoRAs selectively merged back in
- ✅ Manual tuning using merge weighting, style balancing, and… trial-by-GPU-failure
Fun fact: 1 out of every 4 training runs failed. So yes, your prompts aren’t the only thing that gets unstable.
- Stylized Precision: Default outputs are vivid, high-detail, and slightly polished — favoring “clean realism” over documentary grit.
- Portrait-Grade Quality: Close-ups are sharp, defined, and don’t need Adetailer.
- Cinematic Depth & Lighting: Reworked light behavior gives more drama per pixel.
- Prompt Simplicity: No trigger words. Just describe what you want. Seriously.
Expect strong highlights, bold lighting, high-contrast renders, and an aesthetic that sits halfway between real-world and a concept art board.
⚠️ It can do raw realism — but it won’t default there. You’ll have to nudge it.
| Feature | Notes |
|---|---|
| Base Model | SDXL 1.0 |
| Resolution | Native 1024x1024 |
| Default Aesthetic | “Rendered real” – clean, vivid, cinematic realism |
| Close-up Performance | Excellent – no Adetailer needed |
| Lighting & Composition | Enhanced depth, strong shadows, balanced contrast |
| Prompting Style | Natural language and/or tag-based — both work well |
| NSFW Output | Unstable — neutered base, partial restoration, use with caution |
These were tested on Forge WebUI, but should translate well to most SDXL interfaces.
-
CFG Scale:
0.3 – 0.8Sweet spot seems to be
0.5 – 0.7depending on lighting and scene complexity -
Sampler:
DPM++ 2M Karras(🥇 recommended)Euler,Euler ADPM++ SDE- Most usual suspects work fine
-
Steps:
15+More steps = more detail, but diminishing returns after 30-ish
-
Face Restoration:
❌ Disable it. Especially for close-ups — it muddies the eyes.
-
Prompting:
- Descriptive sentence-style prompts are highly effective
- Tag-style also works, especially for concept art or fantasy styles
Typhoon SDXL works best for:
- Portraits and close-ups
- Cinematic compositions
- Soft realism and fantasy realism
- Stylized concept art and clean render-style scenes
You’ll get bold visuals with depth, rich lighting, and smooth but defined surface detail.
All images generated natively using Typhoon SDXL without external LoRAs, inpainting, or fixers. What you see is pure model output.
 |  |  |  |  |
 |  |  |  |  |
Building Typhoon SDXL was not a straight line. The process involved:
- Base Checkpoint Training — Starting from clean SDXL 1.0
- Targeted LoRA Development — Custom LoRAs were trained for lighting, facial detail, anatomy, and style
- Manual Merge Tuning — Each LoRA was merged back via experimental merge weights + scaling
- Hand Calibration — This wasn’t a ComfyUI pipeline — merge order and value tweaks were done manually.
Also developed during the process:
LoRA Strength Analyzer— Get math-backed advice on LoRA merge weightsLoRA Epoch Analyzer— Find which epoch actually did the work
Both tools are in alpha and being wrapped into a Gradio UI soon™.
- NSFW output is very inconsistent — some recovery work has been done, but YMMV
- Default aesthetic leans “stylized-real” — gritty realism must be forced
- High-gloss lighting can over-emphasize skin or fabric unless balanced
- Like all SDXL models, heavy prompting can still cause drift if poorly structured
This model is released under a modified CreativeML Open RAIL-M license:
- ✅ Personal, local use is encouraged.
- ❌ Do not merge with other models — Typhoon’s internal balance will break.
- ❌ Do not upload to public generation sites, bots, or APIs.
Basically: use it, don’t abuse it. Keep the storm local.
- Base: Stable Diffusion XL 1.0 (
stabilityai/stable-diffusion-xl-base-1.0) - VAE: stabilityai/sdxl-vae
- Custom Tools: Raxephion/loRA-Strength-Analyser, Raxephion/loRA-Epoch-Analyser
- Hugging Face Repo: Raxephion/Typhoon-SDXL-V1
If you like what I’m building — Typhoon, Tornado, tools, and controlled chaos — and want to help keep the GPUs humming:
It helps cover compute, coffee, and catastrophic merge experiments.