<h1 id="filigree-anima">Filigree-Anima</h1><p>A full finetune of <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima">Anima Base v1.0</a> by CircleStone Labs, trained in two stages on a hand-curated set of pixiv illustrations selected for ornate costume detail, jewelry, embroidery, and rendered (semi-realistic) shading. Stage two adds a focused polish pass on face and eye rendering.</p><p>Compared to base Anima, Filigree-Anima renders noticeably richer accessory and fabric detail — embroidery, gold trim, layered ornamentation — while keeping coherent composition and clean linework.</p><h3></h3><p><strong>V2.0 RECOMMENDED</strong></p><h2 id="how-to-prompt-it">How to prompt it</h2><p>Filigree-Anima reads <strong>both</strong> Danbooru-style tags and natural language, and handles long, descriptive natural-language prompts especially well — it was trained on hybrid tag + prose captions, so full sentences describing lighting, material, and rendering style carry real weight, not just keyword-style tags.</p><p><strong>Short prompts can underperform.</strong> A handful of tags alone will undersell it — give it real sentences to work with, especially for lighting and material description.</p><h3 id="example-natural-language-(recommended-style)">Example — natural language (recommended style)</h3><pre><code>A detailed anime illustration of Sparkle from Honkai Star Rail lying sprawled 
on a wooden stage floor, viewed from a high angle looking straight down at 
her, her body angled diagonally across the frame with one arm flung above 
her head and the other bent with her hand raised in a playful gesture near 
her face, one knee drawn up and the other leg extended. Her head is tipped 
back toward the camera with a mischievous teasing expression — eyes 
half-lidded with a wide knowing grin and a faint blush.

She has long pink hair with dark inner strands worn in high twintails tied 
with large ribbons, spilling out and fanning across the floor around her 
head. Her eyes are bright magenta-pink with unusual patterned pupils. A 
white theatrical mask with painted markings sits tilted beside her head, 
and she wears a black and pink outfit with a high collar, gold trim, a 
large decorative bow at the chest, dark detached sleeves with pale cuffs, 
and a short pleated skirt.

Around her, dark polished stage boards with scattered confetti and drifting 
paper streamers, a faint spotlight pooling around her.

Magenta, black and cream palette with warm gold accents, dramatic overhead 
theatrical lighting with strong rim light along her hair and shoulders, 
glowing highlights in her eyes, floating light particles, shallow depth of 
field with soft bokeh at the edges, soft cel shading with crisp clean 
lineart, high angle perspective, full body composition, vertical framing, 
high detail.
</code></pre><h3 id="example-tag-forward-style-(also-works)">Example — tag-forward style (also works)</h3><pre><code>1girl, sparkle (honkai: star rail), solo, 
high angle, top-down view, lying, lying on back, sprawled, diagonal framing, full body, vertical framing, 
arm above head, hand near face, playful gesture, one knee up, extended leg, head tilted back, 
looking at viewer, mischievous, teasing expression, half-lidded eyes, wide grin, smirking, blush, 
magenta eyes, patterned pupils, glowing eyes, 
long hair, pink hair, twintails, high twintails, dark inner hair, hair ribbon, hair spread out, 
theatrical mask, kitsune mask, mask beside head, 
black and pink dress, high collar, gold trim, chest bow, detached sleeves, cuffs, short pleated skirt, 
wooden floor, stage floor, dark polished floor, confetti, paper streamers, spotlight, 
magenta and cream palette, gold accents, dramatic lighting, overhead light, rim lighting, glowing light particles, depth of field, bokeh, soft cel shading, crisp lineart, highly detailed
</code></pre><p>Tags carry the concrete content; a short trailing sentence or two on lighting/rendering still helps even in this style.</p><h2 id="no-trigger-word">No trigger word</h2><p>This is a full finetune, not a LoRA — there's no activation token required. It responds to any prompt the base model would; the difference is in the general rendering, not a specific keyword.</p><h2 id="recommended-settings">Recommended settings</h2><p><code>| Setting | Value |</code></p><p><code>|---|---|</code></p><p><code>| Sampler | Euler a |</code></p><p><code>| Scheduler | Simple |</code></p><p><code>| CFG Scale | 5.5 |</code></p><p><code>| Steps | 30-35 |</code></p><p><code>| (Tested) Resolutions | 832x1216, 1064x1568, 1216x1792 |</code></p><h2 id="notes">Notes</h2><ul><li><p>Trained on <code>anima-base-v1.0</code>, not the aesthetic-tuned variant, for maximum headroom.</p></li><li><p>Works well with a hires-fix / tiled upscale pass for additional fine detail on fabric and accessories.</p></li><li><p>Subject to the CircleStone Labs Non-Commercial license, inherited from base Anima.</p></li><li><p>Resolutions can go up to 1440p and native 1440p too. Recommended to use high-res fix.</p></li></ul><p></p>

Files with SHA256: 04e62d47fea44c6b7db04f2db743d09b2a47fbfd6c62de40873a810f4203e388