Aria Solari is my original character: a 23-year-old barista in Valencia, with copper-red hair in two high messy space buns, light blue-green eyes, winged eyeliner and deep matte burgundy-black lipstick. She reapplies the lipstick before she walks into any building. Her Samoyed, Zephyr, goes everywhere with her.
This is my third public Krea 2 character LoRA, after <a href="https://civarchive.com/models/3005329" rel="ugc">Shawy</a> and <a href="https://civarchive.com/models/3007185" rel="ugc">Zara</a>. The trigger is <code>ariasolari</code>. It sits between the two: like Shawy's, the trigger brings her hair, eyes and lips with no description, and like Zara's, at 0.8 it wants the house style sentence. How she found Zephyr, in the rain: <a href="https://civarchive.com/articles/27804" rel="ugc">Rainy Day Rescue</a>. Her profile: <a href="https://civarchive.com/articles/27773" rel="ugc">OC Profile: Aria Solari</a>.
HOW TO USE
<ol>
<li>Load it in ComfyUI with <code>LoraLoaderModelOnly</code> at strength 0.8.</li>
<li>Open the prompt with <code>A painterly anime illustration in a soft cel-shaded semi-realistic register.</code> At 0.8, a prompt without it can come out photoreal.</li>
<li>Put <code>ariasolari</code> where her name would go. It brings her copper-red buns, blue-green eyes and burgundy-black lips, so you do not need to describe them.</li>
<li>Bare prompt? Use 1.0. If the prompt is the trigger plus a scene and nothing else, raise the strength to 1.0. At 1.0 even a bare prompt comes out drawn.</li>
<li>Her things were captioned in training, so call them by the same words: <code>a small hoop earring</code>, <code>a big fluffy white Samoyed dog</code> for Zephyr, <code>an acoustic guitar</code>, <code>white roller skates</code>.</li>
<li>The rest is normal Krea 2 prose: scene, outfit, pose, light.</li>
</ol>
Minimal prompt, tested at 0.8 (896x1152, seed 5555555). It has the house sentence, a scene and the trigger, and nothing about her hair colour, her eyes or her lips. It is the first showcase image:
<pre><code>A painterly anime illustration in a soft cel-shaded semi-realistic register, bright midday palette of orange, leaf green and warm ochre. Drawn in clean cel-shaded anime line art with flat colour areas and simple shading, drawn rather than photographed. A plaza in Valencia's old town at midday. A row of orange trees heavy with ripe fruit stands along the edge of the square, ochre and pale blue house fronts with iron balconies rise behind them, and dappled sunlight falls through the leaves onto the paving.

Framing: Eye level. The camera is at her eye height, looking straight at her. She stands under the nearest orange tree in the middle of the frame, cut at her hips by the bottom edge, its low branches spreading across the top of the frame above her buns.

ariasolari stands under the orange tree, wearing a cropped mustard-yellow sporty hoodie with the sleeves pushed to her elbows and high-waisted denim shorts, with a small hoop earring. A playful closed-lip smirk pulls at one corner of her mouth, as if the orange is already hers. She is alone in the frame, no other people.

Posture and light: She stands with her weight on one hip, turned a little to one side. One hand cups a ripe orange on a low branch at the height of her shoulder, and her other hand holds a white paper cup of coffee at her side. Dappled midday sunlight through the leaves falls across her face, her buns and the mustard-yellow hoodie, and the house fronts behind her soften into ochre and blue. Her lids are lowered a little and her gaze rests on the lens. Gentle depth of field.</code></pre>
IF SOMETHING DRIFTS
Brown hair, pale lips, a face that reads younger than 23. Raise the strength to 1.0. In testing, a bare prompt in a white karate gi did all three at 0.8, on every checkpoint, and 1.0 fixed it.
She goes photoreal. The house sentence is missing. Add it, or use 1.0. The café portrait in the showcase is what a bare prompt at 0.8 looks like: still her buns and her lips, but a photograph.
RECOMMENDED SETTINGS
<ul>
<li>Checkpoint: <a href="https://civarchive.com/models/958009?modelVersionId=3139241" rel="ugc">RedCraft 3.0 (Krea 2)</a>. The only one I tested.</li>
<li>8 steps, CFG 1, sampler euler, scheduler simple.</li>
<li>About 1 MP: 896x1152 portrait, 1152x896 landscape or 1024x1024 square. All three were in the test.</li>
<li>Strength 0.8 with the house sentence, 1.0 when the prompt is bare. Below 0.8 is untested.</li>
<li>Other checkpoints and other UIs are untested. If you try it on plain Krea 2 Turbo or somewhere else, tell me how it went.</li>
</ul>
TESTED SCENES
Before release I ran 17 test prompts at 0.8 on every saved checkpoint from step 1000 to 2500, next to the same prompts with no LoRA. The scenes:
<ul>
<li>a café counter, in a green apron</li>
<li>a portrait in a café at night</li>
<li>roller skates on a seaside promenade, seen from behind</li>
<li>nude on the edge of a bed, a sheet across her lap</li>
<li>a white karate gi in a dojo at dawn</li>
<li>an acoustic guitar on a rooftop terrace at sunset</li>
<li>black lace lingerie on the edge of a bed</li>
<li>a walk on the beach with Zephyr</li>
</ul>
Seven of the prompts are short face prompts with the trigger alone, which is the hard case. On step 1750, six of the seven kept both the copper buns and the dark lips. The seventh is the gi, above.
Then three styles on steps 1750 and 2000, each with a full prompt and a short one. The style itself held in all three:
<ul>
<li>my house style, clean cel-shaded painterly anime</li>
<li>the style line from her character sheet, with warm Mediterranean daylight</li>
<li>a poster in the style of Persona 5, red, black and white with ransom-note lettering. It got the palette, the ransom-note ARIA, the copper buns and the dark lips.</li>
</ul>
Without the trigger the LoRA stays out of the way. A control prompt for a blonde woman, on two seeds, matched the renders with no LoRA. A control for a different redhead, with long loose hair, kept her hair loose, with no buns and no dark lips.
LIMITS
<ul>
<li>Not tested: two characters in one image, crowds, strengths below 0.8, other checkpoints.</li>
<li>Explicit frames are not a target, and are not tested.</li>
<li>Nude frames come out very slim, with visible ribs.</li>
<li>Nails come out dark in most frames, and red-brown in some short-prompt frames.</li>
<li>One style test went soft: the character-sheet style with the trigger alone gave her lighter, strawberry hair and natural lips.</li>
</ul>
MAKER'S NOTE
Trained on Krea 2 Raw with ai-toolkit (arch krea2), rank 32, batch 1, lr 1e-4, 1024 px, 32 images. The file is step 1750 of a 2500-step run, about 54 epochs. The dataset is synthetic: 32 images of her rendered on Krea 2 from her written description, 25 made for this set and 7 kept from earlier renders of her. 20 dressed, 7 lingerie, 5 nude.
No v2 this time. Shawy and Zara both needed a second training. Aria's first one was fine, which has not happened to me before.
I picked step 1750 over the final 2500. The lipstick was the last thing LoRA learned: at step 1000 it was red everywhere, at 1500 dark only in some styles, at 1750 dark everywhere. She reapplies it before every building, so I respect the priorities. Past 1750 the bare prompts start losing her buns and her lips again, and by 2500 some of them give her brown hair and red or pale lips. The training loss stayed flat the whole run, so the curve shows none of this. Only the pictures do.
Aria is my own character. Use the LoRA for your own non-commercial work, and credit this page when you share it.

Unseen Art Aria Solari (Krea 2)

v1, step 1750. The first training was good enough, so there is no v2.
I picked step 1750 over the final 2500. Her lipstick only turns dark everywhere at 1750, and past it the bare prompts start losing her buns and her lips again, while the loss stays flat.
Strength 0.8, with the house sentence from the model page.
Bare prompt? Use 1.0. That is the trigger plus a scene and nothing else. At 0.8 a bare prompt can come out photoreal. At 1.0 it comes out drawn.
Tested only on RedCraft (Krea 2), 8 steps, CFG 1, in ComfyUI.

Files with SHA256: 853b1c72d0df47099ce1a54297ba3ccc6173785272ec78d9af90c69a78312685