grafelem style

Style LoRA based on illustrations by grafelem.

Loose painterly anime: brush-stroke hatching on clothes and hair, rough painted backgrounds, sharp ink lines on top.

The whole picture is drawn in one hand — and the rough strokes also hide the small mistakes the base model usually makes in hands, gear and busy backgrounds.

Works on anything: single characters, two- and three-person scenes, landscapes with no people, animals, still life.

This is a style, not a character — name the character you actually want.

Not an official model by the artist.

Base model: Anima (anima-base-v1.0)

Trigger: @grfl. (put it at the very start of the prompt, then regular Danbooru tags)

No style tags needed.

LoRA strength: 1.0 (0.8–1.2)

Example prompt

@grfl. safe, masterpiece, best quality, score_7, 2girls, frieren, fern (sousou no frieren), purple hair, sitting at tavern table, eating, bread, soup, warm light, smile

Recommended settings

• Sampler: DPM++ 2M SDE, scheduler simple, 25 steps, CFG 5

• Hires fix: 1.5×, lanczos, 20 steps, denoise 0.35 — recommended: it keeps the hatching and adds crisp line detail.

Negative prompt:

worst quality, low quality, score_1, score_2, score_3, jpeg artifacts, logo, watermark, signature, twitter username, bad anatomy, bad hands, missing finger, extra digits, fewer digits, text

Tips

• Known characters work as usual (hatsune miku, roxy migurdia, makima (chainsaw man)). If a second character comes out wrong, add their hair colour — e.g. Fern needs purple hair.

• Don't add extra aesthetic tags like score_8 or aesthetic — they push Anima towards a smooth polished render and wash out the strokes.

• Strength above 1.2 is not needed and starts to override the composition.

────────────────────────────────────────────────────────────

Training

105 images, signatures and watermarks removed · short captions: trigger + people count, framing, hair colour, a few key items and character names — nothing describes the style or colours, so the LoRA carries all of it

No quality tags in captions: the model keeps its own sense of quality and the style is layered on top

Anima LoRA, rank 32 / alpha 32, incl. AdaLN modulation · AdamW8bit, LR 1e-4, constant with 5% warmup · 12 epochs, 1260 steps, batch 1 · 1024 res, buckets 512–1536 · sigmoid timesteps, scale 1.3

Epoch 12 picked by side-by-side, image-by-image comparison across 18 scenes and several seeds