Atem is a dark, edgy, atmospheric checkpoint series, built around deliberate roughness. Heavy contrast, domineering characters, a drive to make your prompts edgier. The goal is output that resists being sanded into a generic anime look.

What this model is good at

- Gothic, horror-adjacent, and atmospheric compositions

- Strong lighting contrast, deep shadows, selective glow

- Characters with presence: dark elegance, quiet menace, mysticism

- Cool color grading (teal, cyan, violet) with selective warm accents (blood red, amber)

What this model isn't good at

- Not built around explicit content. It can produce it, but I didn't tune it for that.

- Not a clean, soft anime look. If you want polish and brightness, this isn't the tool.

Recommended settings

- Sampler: Euler Negative Dy, Restart, DPM++ 2M Karras. (Don't use Hiresfix for Euler Negative, unless you know what you are doing.)

- CFG: 4-6 (higher CFG introduces "Splash Art Styles" occasionally, where the image is framed by white)

- Steps: 30 for Euler Negative Dy. Feel free to use other steps for Restart and DPM++ 2M Karras

- Resolution: Standard SDXL resolutions. Prefers 832x1216, 1024x1024, or 1216x832 for native performance.

Prompt guidance

-Atem works best with minimal negative prompts. A long negative neutralizes exactly the qualities this merge is built for. Trust the base and prompt sparingly against it. The recommended negative prompt:

worst quality, (low quality, normal quality:1.3)

-For atmosphere: lean into scene description (lighting, mood, environment) rather than quality tokens. Atem responds strongly to environmental cues.

-For characters: describe specific aesthetic details (material, texture, expression) rather than generic quality words. "Black silk, torn at the edges" outperforms "high quality clothing."

Merge composition

Built on Illustrious XL 2.0 base, merged with 2.5D style models. Added dark fantasy LoRA, then re-merged with a few anime models to stabilise excessive artifacts.

The point is to introduce structural noise that standard Illustrious merges sand away, while keeping the products stable.

License and use

Feel free to use outputs commercially. Merging into your own work is welcome credits are appreciated but not required.

No training data restrictions beyond what the base license carries.

If you build something interesting with this, I'd love to see it. Tag me on Twitter (@TridIsntAName) or drop a link in my dms!