🌪️ Typhoon (Stable Diffusion 1.5 Edition)

V1 and V2 – from experimental mess to high-res chaos.

"What began as Tornado V3 spun out into its own category-five situation."

“Against better judgment and diminishing funds, here's the SD1.5 adaptation of Typhoon.”

Typhoon (SD1.5) is the smaller sibling of my flagship SDXL model, brought to you through equal parts curiosity, stubbornness, and significantly less dataset-induced remorse.

(Typhoon XL: https://civitai.com/models/1610717/typhoon)

📌 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.

This showcase uses Native Output Samples: All sample images were generated without any LoRA or external modifiers. What you see is core model capability.

📦 What is this?

This is a custom SD1.5 model trained in phases: initial checkpoint finetuning followed by training and merging purpose-built LoRAs — all done by hand, the hard way. Think of it as a Frankenstein’s monster stitched together with LoRA gauze and learning rate duct tape.

The base is 'v1-5-pruned-emaonly', since the original runwayml/stable-diffusion-v1-5 wasn’t available anymore. Choosing a base model was a nightmare — most were either fine-tuned to oblivion or crippled beyond recognition. This one barely survived.

🔧 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. That said, it can do raw-realism—you just have to work a bit harder for those results with prompting, and steer away from the high-gloss trigger phrases.

📚 Training Notes

The dataset has been fully overhauled and upscaled to 768px, correcting the sins of V1. Additional material was also added to improve compositional variety and general anatomy. You may now throw fewer chairs when generating full-body images.

LoRAs used in the merge were trained on 512x512 in V1, but are now backed by improved datasets.

All sample images were generated without any LoRAs.

Resolution tip: While you can generate at 512 x 512 or 512x768 or 576x768, expect some anatomical weirdness at times. I recommend starting with 512 x 768, especially if you're not using hires fix, but really, it can make a big difference.

🧠 Under the Hood

Typhoon (SD1.5) was created using a hybrid process:

• Initial base checkpoint fine-tuning

• A suite of project-specific LoRAs, trained and manually merged into the base

• Careful balancing of merge strengths and weights, mostly through trial and error

To assist with this process, I built two open-source tools:

📊 LoRA Strength Analyzer
Helps evaluate LoRA merges using perceptual image metrics like SSIM and BRISQUE.

📈 LoRA Epoch Analyzer
Compares different LoRA training epochs to identify optimal checkpoints using visual metrics.

⚠️ 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: Vastly improved from V1 thanks to the dataset fix, but edge-case weirdness may still occur. Use Hires Fix whenever possible.

Natural Language Prompts: Doesn’t like them. Stick to short, structured prompts for best results.

🧠 Prompting Tips

This model does not like long, flowery prompts. It's not Flux. You don't woo it. It prefers concise, tag-like prompting. Stocatto. Likely because the training sets were tag-heavy. Again, my boo-boo, or booru. But seriously, you’ll get the best results keeping things short and specific.

⚙️ Recommended Settings

• Resolution: 512x640 or 576x768 for base gen

• Sampler: DPM++ 2M Karras, Euler, Euler A — the usual suspects

• CFG: 0.3 – 0.8 for most images

• Steps: 20 - 40 is usually good. Increase or decrease by -+5 steps if/when results are bad

• Hires Fix: Highly recommended

• Denoising: 0.5 - 0.7 (0.5 = softer image)

• Upscale: 1.5 - 1.8 x (Latent) = Higher = more artifacts

• Hires CFG: 7 (I set hires fix cfg to the same value as my base cfg — I know, OCD...)

⚠️ Do not use ADetailer or face restoration for closeups — this model already renders eyes and faces well. Use only if the model struggles in edge cases.

💡 VAE Note

I tested one of those older .bin VAE files during development — it dulled the colors and generally made things look like they took a detour through Berlin in October. Dull, dreary, and ...you get the picture.

Typhoon works best either:

• With no VAE loaded at all, or
• With the official stabilityai/sd-vae-ft-ema.

If your results look washed out or lifeless, the VAE is probably to blame.

🔒 License & Usage

Licensed under CreativeML OpenRAIL-M.

• Do not merge this model into others.

• Do not use this model on third-party generation services.

• You may use it for private, personal generations.

• If you want to use it in a commercial or hosted context, ask me first.

Merges or reuploads will break the aesthetic and completely ruin the careful LoRA balancing. Let’s not do that.

🚧 Known Issues

• Some anatomical glitches may still occur at narrow resolutions.

• Not optimized for natural language prompts.

• Improved dataset resolution in V2 resolves most of the full-body distortion issues from V1, but edge cases remain.

🌩️ ☕ 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.