Fever Cel (熱病のセル画)

Pen and graphite horror anime for Krea 2

Fever Cel draws your scene the way an underground horror anime from the early 2000s might have: very fine black pen lines, graphite shading, dense cross-hatching and the grit of a cheap photocopy, with toner specks and halftone grays. It's black and white by default, and it likes to make things a little wrong. Masks, uniforms, moths and centipedes, pale creatures with too many arms, long black hair spilling out of a doorway.

There is no trigger word, just describe the scene. Two short phrases, explained below, let you add a single accent color or some hand-drawn glitches when you want them.

Settings

• Krea 2 Turbo, 8 steps, CFG 1, Euler / Simple, 832×1248

• LoRA strength 1.0 (0.8 if you want lighter, cleaner lines and a bit less grit)

• Hires pass: ×1.5 latent upscale, 4 steps, denoise 0.4, which gives 1248×1872

• Detail Slider at 0.4 and Aesthetic Masterpiece at 0.3 for a little extra polish (optional)

Strength

The first comparison shows one scene on one seed three times:

Krea 2 Turbo on its own, then Fever Cel at 0.5, then at 1.0. At 0.5 the image is already fully drawn, with pen lines, hatching and no color, but it's calmer and cleaner, the face looks softer and more realistic, and there are fewer stains. At 1.0 you get the whole look: deeper blacks, heavier hatching, toner stains and specks, and a more stylized anime face. The composition stays the same at every strength, so you can tune it without losing your scene.

Adding an accent color

Put with red ink accents (or another color) at the end of your prompt. The LoRA then overprints loose ink on a few details, a little out of register and sometimes dripping, and adds a thin offset outline along some edges, like color fringing. The rest stays black and white.

It was trained on eight colors, and those are the reliable ones: red, orange, yellow, green, teal, blue, violet and pink. Other names are hit and miss ("cyan" came out violet in my tests), so pick the closest of the eight. Leaving the phrase out gives you pure black and white, which is the intended default.

The second comparison shows no phrase, red and teal on the same seed. The scene barely changes; the color mostly shows up as an outline around the figure and a few objects. On a busy scene the effect is subtle. Red is the strongest, while teal, blue and violet are quieter. It gets much bolder when the scene has things it can paint, like lanterns, a kimono, a mask or curtains, as you can see in the gallery.

Adding glitches

Put with glitch artifacts at the end of the prompt, after the color phrase if you're using one. You'll get a few stair-stepped pixel blocks and shifted slices, drawn in graphite like everything else. They stay small and graphic and don't wreck the drawing.

Use that exact wording. In the third comparison, "with heavy glitch artifacts" produced no glitches at all on the same seed. The LoRA learned that one phrase, so adding words like "heavy" or "subtle" breaks it rather than scaling it.

An example prompt

The scene comes first and then the switches:

A nurse in a fitted white uniform and a white kitsune fox mask stands in a hospital corridor at night, the many pale arms of a creature fanning out behind her, moths everywhere, with red ink accents, with glitch artifacts.

Prompting tips

Describe what's in the picture, not the style. The LoRA was trained without style words, so you don't need "pen", "graphite", "manga" or "black and white". And "anime" can actually pull it toward a cleaner, glossier look.

It looks best with the kind of content it was built around:

• porcelain and kitsune masks

• uniforms: nurse, police, maid, office, black suits

• pale moths, centipedes and beetles

• long black hair and pale many-armed beings

• flooded rooms, moons, empty corridors at night

Concrete details work much better than mood words.

"Moths on the windows, a centipede on the seat" works better than "creepy".

It can draw faces well, but it is made to be used with masks, hair or a turned head. If you want a face, write "looking straight at the viewer" or "her face clearly visible".

Bright, cheerful, photo-style prompts push back. Something like "a woman in a summer dress at a picnic, smiling" can stay photographic. Give it something to draw, a night setting or one strange detail, or raise the strength. Newspapers, signs and screens tend to fill up with fake Japanese text, which a no-text line keeps down.

Works well with those LoRAs

• Detail Slider at 0.4 to 0.6 makes the hatching finer and adds texture.

• Aesthetic Masterpiece at 0.3 adds a light polish. Keep it low or things turn glossy.

• TextFusion Refusal-Reduction at 0.8 to 0.9 makes Krea follow the darker, stranger parts of a prompt more literally.

Good to know

If color shows up when you didn't ask for it, check the scene for color words; "a red umbrella" will be red. Large flat areas pick up toner specks and grit, which is part of the look, and the Detail Slider or a hires pass makes it finer.

Thanks for checking out my LoRA and I hope you find it helpful in your artworks.
Feel free to leave some feedback or let me know what could be improved. 🫶