CivArchive
    Sabikaze - grey-teal dry-brush anime style (Anima) - v1
    NSFW
    Preview 146470621
    Preview 146470622
    Preview 146470624
    Preview 146470644
    Preview 146470632
    Preview 146470655

    Sabikaze · Anima style LoRA

    Quiet grey-teal palette · one warm accent · dry-brush ink · light grain

    Sabikaze turns clean anime renders into muted, painterly illustrations: a desaturated grey-blue base, a single small spot of warm colour, loose confident ink and flat-brush strokes, and a figure that sits inside a real place.

    Quick start

    @z3-sabikaze, 1girl, solo, adult woman, short hair, black hair, bob cut, high collar, coat, looking to the side, upper body, outdoors, harbor, cloudy sky
    • Trigger: @z3-sabikaze at the start of the prompt

    • Strength: 1.0

    • Base model: anima-base-v1.0

    • Sampler / scheduler: er_sde / simple

    • Steps: 40 · CFG: 4.5 · Shift: 3

    • Resolution: around 1 MP (896×1152, 768×1344, 1216×832, 1024×1024)

    • Negative: lowres, worst quality, low quality, bad anatomy, bad hands, jpeg artifacts, signature, watermark, text, logo

    Works best for

    • Figures placed in an environment: rainy streets, harbors, piers, quiet rooms, fields at dusk

    • Upper-body and full-body shots, men and women, older characters

    • Interiors at night, where the grey-teal base and one warm lamp come through clearly

    Known limits

    • Close-ups on a plain or white background stay close to clean base-model anime. Give the prompt a setting.

    • Bright outdoor scenery tends to keep a saturated blue sky.

    • Only strength 1.0 was tested for this version.

    Training

    • Data: 57 images I generated myself, hand-picked to match one coherent look

    • Run: 12 epochs, 2052 steps; this release is epoch 8, the closest match to the target look

    • Network: LoRA dim 32, alpha 32, lr 5e-5 constant, AdamW8bit, bf16, sd-scripts

    Changelog

    • v2 - first public release. Narrower dataset and a stronger learning rate than the unreleased v1, which stayed too close to clean saturated anime.

    Description

    Earlier, weaker version, uploaded for comparison with v2. Trained on 143 images at lr 2e-5 for 12 epochs (1716 steps); this is the final checkpoint. The style comes through on figures in a setting, but close-ups stay close to clean, saturated base-model anime. The showcase uses the same prompts and seeds as the v2 showcase. Use v2 for the intended look.

    FAQ

    Details

    Downloads
    17
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/22/2026
    Updated
    11/4/2026
    Deleted
    -
    Trigger Words:
    @z3-sabikaze

    Files

    sabikaze_anima_v1.safetensors

    Mirrors