Typhoon XL
Details
Download Files
About this version
Model description
📌 Heads-up: CFG scale recommendations was a typo - corrected
📌 TL;DR / Heads-up:
Typhoon is a stylized model. It’s intentionally biased toward enhanced realism—cinematic lighting, clean detail, and a signature “rendered-real” aesthetic. It’s capable of full photorealism (think smartphone-tier portraits, polaroids, film grain, etc.), but raw realism requires more prompting effort. By default, expect vivid detail and that elevated, slightly surreal polish—part of what gives Typhoon its unique charm.
📌 Update (02/07/2025): V2 is now live!
Typhoon V2 builds on the visual foundation of V1 with improved prompt adherence, better anatomical precision, enhanced material realism, and tighter control over lighting and surface detail. The changes may seem subtle at first glance—but this was a surgical upgrade: intentionally focused, highly specific, and technically complex. Nothing was overhauled. Only what needed fixing was fixed.
This showcase uses Native Output Samples:
All sample images were generated without any LoRA or external modifiers. What you see is core model capability.
🌪️ Typhoon – A High-Fidelity SDXL Model for Precision and Artistry
Typhoon is not a mere continuation—it is a deliberate departure from its predecessor, Tornado. Originally envisioned as Tornado v3, the project evolved into a standalone model with a unique visual identity. The result is a high-performing SDXL model built for clarity, depth, and cinematic composition, delivering richly detailed and stylistically versatile results across a broad range of subjects.
Development required extensive dataset curation and GPU-rented training runs. With limited hardware access, the process was both resource-intensive and costly. On average, one out of every four training runs failed, making meaningful progress difficult and unpredictable. Typhoon is the product of that persistence—carefully constructed through trial, refinement, and iteration.
🔧 Model Characteristics
Portrait-Grade Detail: Typhoon excels at clean, high-resolution close-up rendering. ADetailer is not required—detail retention is natively strong.
Dynamic Lighting & Depth: With re-engineered light behavior, Typhoon offers improved cinematic contrast and enhanced spatial composition.
Prompt Simplicity: No trigger words are needed—results are natural, responsive, and consistently aligned with prompt intent.
Typhoon is stylized by nature. Its default output is vivid, detailed, and lightly rendered—favoring a “clean realism” or “photoreal-adjacent” look over pure, raw realism. Expect bold lighting, sharply defined features, and that signature rendered-polish look baked right in. It’s excellent at portraits, beauty shots, and concept art, but it won’t default to gritty or documentary-style realism unless prompted carefully.
Typhoon V2 builds on all of this—while also improving response to prompt phrasing. Where V1 sometimes resisted certain pose or orientation cues, V2 is much more obedient. (The sample comparison includes a prompt using “back to viewer”—which V2 handles with precision.) Eye rendering has also been noticeably improved at medium range: while V1 already performed well at close-up distances, it could falter slightly at mid-range. V2 corrects this, delivering sharper, more expressive eye detail even in wider compositions. Material rendering and environmental separation have also been enhanced without affecting the overall aesthetic or tone. It’s still Typhoon—just... tighter.
⚙️ Recommended Settings
The following apply to both V1 and V2, as settings are fully compatible:
(Note: On Forge WebUI I had to disable face restoration otherwise the eyes are smudged, especially for close-ups. This model handles close-ups very well straight out-the-box.)
CFG Scale: 3.0 – 8.0 (sweet spot: 5.0 – 7.0 depending on style)
Steps: 15+ (more steps can enhance complexity)
Samplers:
DPM++ 2M Karras (recommended)
Euler / Euler A
DPM++ SDE
Other “usual suspects” may also yield excellent results
ADetailer: Not recommended for close-ups—can cause eye smudging
Prompting: Works better with descriptive sentence structure
Style Fit: Excellent for portraits, cinematic renders, fantasy art, soft realism, and stylized concept work
🛠️ Development Process
Typhoon's construction involved a hybrid approach:
Initial Training: Began with traditional checkpoint training to establish the base model.
Targeted LoRA Integration: A suite of LoRAs was trained specifically for Typhoon. These were then selectively merged into the base model.
Manual Tuning: The merging process itself was trial-and-error—requiring repeated experimentation with merge weights, strength values, and order of operations to maintain balance and prevent style bleed.
This hands-on refinement led to the creation of two custom tools:
(I am in the process of developing a Gradio UI for these two scripts to make it easier for non-engineers and those who don't like fiddling with code and command prompts)
LoRA Strength Analyser – a Python script that analyzes LoRA influence by strength and offers math-guided recommendations.
LoRA Epoch Analyser – provides insights into which training epochs produce the most consistent outputs.
Both tools are currently in alpha and are freely available via GitHub.
⚠️ Limitations
Typhoon is not fully consistent in NSFW generation. The base model was heavily neutered in this regard, and while V2 has improved reliability compared to V1, it’s still not perfect. When it works, it can impress—but results may still vary depending on the prompt and composition.
🔄 Ongoing Updates
Typhoon is an evolving project. V2 is the latest milestone, but not the last.
Updates will be released as development continues. User feedback and extended testing will continue to shape improvements over time.
📜 Usage Policy & Disclaimer
Typhoon is intended for personal and creative use only.
Please do not upload or deploy this model on third-party generation services, and do not merge it with other models. Typhoon is designed as a tightly calibrated system, and altering it may significantly degrade its visual balance and intentional fidelity.
🌩️☕ 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.