magful — Krea 2 Thick-Paint Style
A thick-paint / impasto style LoRA for Krea 2 with a built-in colour strategy — not just a texture filter. It doesn't lock your composition and doesn't override your art style.
Trigger word: magful — OPTIONAL. The LoRA is always-on; adding it makes no systematic difference.
Recommended: strength 1.0 | cfg 1.0 (daily) or 1.5 (punchier)
⚠️ READ FIRST — WANT AN EMPTY SCENE / SCENERY ONLY?
This is where people get tripped up. It absolutely can do empty scenes — short prompts just can't hold back its default configuration.
38 empty-scene renders (6 scene types × 3 seeds):
- Short prompt (~10 words) + poster-art prefix: characters appeared 18/18
- Short + "no humans" at the very front: 14/14 (even with no extra words at all)
- Medium detail (~47 words): 0/2 — clean
- Long prompt (200+ words, per-object spec): 0/2 — clean
- Still-life subject (no poster prefix): 0/6 — clean
- "dark fantasy aesthetic" tone: 0/4 — clean
WHAT TO DO
1) Write enough objects and space — counts, positions, materials, light direction.
Same empty classroom: 10 words → 4-7 students fill the desks; 47 words (blackboard, window direction, dust, chair states) → completely empty; 234 words (five rows of thirty desks, clock at 4:10, aluminium window frames, long floor shadows) → still empty, and composition adherence is excellent.
2) Avoid the poster-art prefix.
"official poster art / cinematic composition" semantically means "a key visual WITH a subject". Same empty classroom, only the prefix changed:
- no prefix / magful / "no humans" first: classroom still gets filled with students
- "Modern high-end anime illustration, official poster art, ..." : even a still life becomes a silver-haired portrait
- "dark fantasy aesthetic": classroom stays completely empty
Moving "no humans" to the front does NOT fix it (4/4 either way).
3) Use cfg 1.5 — cuts the head-count a lot (5-6 people → 1), but doesn't remove it.
One line: empty scenes come from writing detail, not from adding negatives.
WHAT IT DOES
Controlled experiment (fixed prompt × 4 seeds × 3 strengths, 36 renders):
LoRA colorfulness = 0.896 × base colorfulness + 3.93 (R² = 0.995, strength 1.0)
Crossover at base ≈ 37.7 — below it the LoRA ADDS colour, above it MUTES.
- Gothic / dark (nun in cathedral): base 15.1 → 18.5 (×1.23, more colour)
- Mid-tone (afternoon street): base 35.8 → 34.5 (×0.96, about the same)
- Saturated (magenta/cyan flower field): base 95.8 → 69.9 (×0.73, muted)
In both directions: distinguishable colours +40% to +174%, luminance micro-detail +3% to +49%, stronger colour peaks (the most saturated 5%), especially in dark scenes.
It doesn't simply "add saturation" or "go grey" — it pulls every starting point toward a "composed but powerful" middle range. Low-chroma images gain colour, high-saturation images gain gradation. Both make the picture read thicker.
STRENGTH & SETTINGS
Strength 1.0 — thickest layering, most restrained chroma (recommended)
Strength 0.8 — middle ground
Strength 0.6 — flattening out
Strength 0.4 — close to base
cfg: 1.0 for daily use, 1.5 for more punch (chroma +5.6% to +21.6%). The colour behaviour keeps the same direction at both.
Sampling: euler / sgm_uniform / 8 steps | CLIP: Qwen3-VL 4B | VAE: qwen_image_vae
CONTROLLABILITY (all backed by A/B tests)
- Explicit hair colour: 13/13 (black / blonde / green / pink / brown / silver / purple)
- Explicit head-count on plain background: 6/6 exact
- Explicit photographic style: 8/8 fully photoreal — the LoRA steps aside
- Six forced media (ink line art / oil / 3D / photo / watercolour / flat vector): 6/6 kept their medium — NOT invasive
Invasiveness measured: with ink line art forced, luminance gradient 0.1152 vs base 0.1154 — a difference of 0.0002. It only nudges inside a medium (mutes vector colour, thickens watercolour strokes). It will not destroy the style you asked for.
Prompt responsiveness: across 6 very different prompts, within-group distance 0.31 (LoRA) vs 0.36 (base) — distinguishing ability preserved, only ~13% homogenisation.
THINGS TO WATCH
- Unspecified hair tends to silver (the base model already prefers this)
- Unspecified outfit tends to school uniform (same-seed test: 6/6 uniforms vs 6/6 casual)
- Scenes implying company may add a second figure — add "solo, only one person in frame"
- Short prompt + poster prefix fills the frame with people — see the first section
RETRACTED CLAIMS
Several earlier conclusions were overturned by later controlled tests — three of them from this description's own earlier drafts. Listed so nobody propagates them:
- "Overrides explicit photographic style" → 8/8 stayed photoreal, the LoRA stepped aside
- "Miscounts people in group scenes" → 6/6 exact with "3girls" on a plain background
- "Obeys empty-scene prompts 4/4" → that test prompt read "solo, 1girl, only one person in frame, a tiny figure" — it asked for a person. Empty scenes were never tested then.
- "Has a colorfulness floor around 28-31" → dark scenes output 15.6-20.8, no floor; the real relation is linear shrinkage
- "+91% colorfulness in dark scenes" → single uncontrolled sample; controlled value is +10% to +20%
- "Lighting words are bound to characters, so they pull people in" → the no-word control also produced people 14/14; the deciding factor is prompt specificity
"Silver hair" is NOT this LoRA's signature — the base model draws silver-haired girls from "1girl" alone.
EXAMPLE PROMPTS
Reusable prefix:
magful, Modern high-end anime illustration, official poster art, cinematic composition, intricate details, sophisticated style, depth of field, sharp focus on subject, masterpiece.
1. 1girl, gothic nun, silver cross, gothic cathedral, heavy rain, dark dramatic lighting
2. 1girl, valkyrie, silver winged helmet, ornate plate armor, holding a spear, divine golden light, clouds and sky, epic battlefield, dramatic backlighting
3. 1girl, wandering swordswoman, black hair in a high ponytail, worn traveler cloak, katana at her side, cherry blossom grove, wind scattering petals, twilight
4. 1girl, young librarian, brown hair with glasses, cream blouse, standing among towering bookshelves, warm lamplight, dust motes, cozy atmosphere
5. 1girl, elf archer, long green hair, pointed ears, dark green cloak with gold embroidery, drawing a wooden bow, ancient mossy forest, fireflies, moonlight
6. 1girl, cosmic priestess, white hair, layered white and gold robe, standing before a giant celestial harp, starfield, floating planets, ornate golden architecture
Negative:
cartoon, chibi, cel shading, flat colors, childlike, photorealistic, 3d render, low quality, bad anatomy, text, watermark
TRAINING
Base: Krea 2 | trainer: ai-toolkit | 2500 steps / 138 epochs | rank 32
Dataset: a self-curated set (mostly own AI-generated works, plus a small number of third-party AI-generated pieces), style-consistent but structurally diverse; the dataset was not retained after training.
License: non-commercial. Because the dataset includes third-party material, please do not use commercially; derivatives should credit the source.
---
Mirrors
- Hugging Face: https://huggingface.co/weiermod/magful-krea2-lora
- GitHub: https://github.com/weiernmod/magful-krea2-lora
中文说明
magful —— Krea 2 厚涂风格 LoRA。触发词 magful 非必须，默认生效。推荐强度 1.0，cfg 1.0（日常）/ 1.5（更艳）。
想做空景？关键在提示词的具体程度，不在题材。38 张空景测试：短句（10 词）+ 海报前缀 → 18/18 出人；加到 47 词 → 0/2；234 词 → 0/2；静物题材 → 0/6；dark fantasy aesthetic 调性 → 0/4。official poster art 这类前缀是主要开关，连静物都会被画成银发女孩，把 no humans 放最前面救不回来。空景靠「写详细」，不靠「加负面」。
效果：LoRA 鲜艳度 = 0.896 × 基座鲜艳度 + 3.93（R²=0.995），交叉点约 37.7 —— 低于它加色（暗调 ×1.23），高于它减色（高饱和 ×0.73）。两个方向都提升色阶（+40~174%）和明暗微细节（+3~49%）。
可控性：发色 13/13、白底人数 6/6、摄影风格 8/8 让位、六种强制画风 6/6 全部守住（墨线风格下梯度只差 0.0002）。
注意：不写发色会银白、不写服装会制服、场景暗示有同伴会多一个人。
许可：非商用。素材含第三方 AI 作品，训练后未保留。