Freestyle - an anime style LoRA for Krea 2.

TRIGGER WORD

freestyle.

Put it at the very start of the prompt, period included. Captions were trained as "freestyle. <natural-language description>", so the trigger only fires reliably in that position. Krea 2 wants natural language - do not feed it Danbooru tags.

WHAT IT DOES

Pulls toward built-out, in-focus backgrounds, richer and more elaborate outfits, and a more covered read of the character. It is the restrained, illustration-heavy end of the pair.

COMPANION LORA

Built as one end of a continuous axis together with Mretsis Style [Krea 2]:

soft bg / simple outfits / higher exposure <-- mretsis ---------- freestyle --> detailed bg / ornate outfits / lower exposure

Recommended pairings:

mretsis 0.6 + freestyle 1.0 - everyday work

mretsis 0.8 + freestyle 0.8 - spicier scenes

Both triggers go at the head of the prompt: mretsis. freestyle. [style block] [scene]

Note that mretsis has the larger marginal effect of the two - it will blur the background even at low weight, so freestyle usually wants to sit higher.

RECOMMENDED SETTINGS

Base model: Krea 2. Tested on RedCraft 3.0 (Krea 2).

8 steps, CFG 1, euler + simple, 1328x1328.

Weight 0.8-1.0. It works alone, but the pairing above is what it was tuned against.

ON STYLE BLOCKS

With this LoRA at strength, a written style block acts globally rather than through any one sentence - ablating single sentences changes almost nothing, but dropping the block entirely shifts exposure, palette and composition noticeably. Keep the whole block rather than trimming it.

The block used throughout testing:

A modern Japanese anime illustration with flat cel shading: hard-edged shadow shapes, clean uniform line weight and crisp lineart, bright limited-palette colors held in flat fields. Painterly impasto appears only as a thin accent layer on hair highlights, fabric folds and cheek shadows - everywhere else the edges stay hard and the color stays flat. Secondary elements keep volume and lighting but drop fine detail. The background is radically simplified into large flat color fields holding only silhouette and light-dark structure, creating a clear three-tier detail hierarchy from character to environment.

TRAINING

288 images, 3 repeats, 1024x1024 with aspect-ratio bucketing, 10 epochs / 2880 steps.

Network dim 32, alpha 32, AdamW8bit, constant LR 1e-4, bf16 with fp8 base.

Captioned in natural language with the trigger word as a prefix.