Aria Maiven: one dataset, Illustrious against Klein 4B

Aria is a stack of small details: red at the roots fading to pink, amber eyes, freckles, a black ring choker, scissors and a comb on a thin chain, flowers over skulls on both forearms, one skull with roses on her left shoulder. A LoRA that gets the hair and drops the rest gives you a redhead.

This test asked one question: does the LoRA hold her with no reference image, on WAI-Illustrious, the checkpoint I run anime on? Whoever downloads it gets a trigger and a prompt box, nothing else.

Dataset and tagger

33 images, 13 of them in her look (oversized band tee off one shoulder, black bike shorts, thigh strap), mixing close-ups, half body and full body across angles and light.

Illustrious reads Danbooru tags, so the captions are tags, opening with ariamaiven. I ran WD14 and pruned every caption by hand. Keep what varies: outfit, pose, framing, background, expression, light. Delete what's fixed about her: red hair, gradient hair, freckles, tattoo, choker and the rest. Whatever the caption names becomes a separate variable, and whatever it leaves out sticks to the trigger. Every tag about age or a youthful body goes too.

The tagger misread her all the time. The scissors-and-comb pendant came back as "cross", "key", "dog tags" and "skull necklace". Freckles became "mole". Amber eyes became "red eyes". It added "1boy" and "artist name" to images with neither. Read every caption before the zip.

Mirror shots stayed out, because the reflection puts the shoulder tattoo on the wrong side. Flip is off for the same reason.

The trigger is ariamaiven because "aria" alone collides with what the checkpoint already knows. Her outfit has its own trigger, ariamaiven outfit, so you can ask for her look or dress her in anything.

Two trainings, same 33 images

Both on Civitai's trainer, 500 Buzz each:

• Illustrious-XL v0.1, with tags. v0.1 because WAI and NoobAI come from it, and a LoRA trained on v1.0 or v2.0 degrades back in that family. 1024, about 1,500 steps, 10 epochs saved, alpha equal to dim.

• Flux.2 Klein 4B, same images, sentence captions. Klein draws closer to her line, so it got a shot.

Before looking at any output I wrote the bar down. The judge is a Gemini vision prompt that grades each feature against her reference from 0 to 10, where 7 means present and correct. The bar: 6.5 or more in 8 of 10 scenes, sleeves, freckles, choker and chain present in most scenes where they'd show, cel-shaded every time, and zero images reading young. One young-looking image fails the whole epoch.

Illustrious was tested on WAI Illustrious with no reference, across ten scenes that included a close-up, the atelier at night in her look, full body barefoot, a back view over the shoulder, a black party dress, a neon street, a pure profile and a ponytail. Klein ran on Civitai's generator.

The numbers

• Illustrious, epoch 8, weight 0.8: Scenes at 6.5+ 1/10; Average 5.80
• Illustrious, epoch 10, weight 1.0: Scenes at 6.5+ 3/10; Average 6.26
• Illustrious, epoch 10, 1.0, plus freckles, tan skin: Scenes at 6.5+ 6/10; Average 6.40
• Klein 4B, epoch 8, weight 1.0: Scenes at 6.5+ 1/9; Average 5.86

Illustrious won, though not like for like: Klein ran 9 scenes at epoch 8, without the two tags that lifted Illustrious. Klein got the line and the hair gradient closer, then dropped the rest. It never learned the outfit (it drew pajamas and a robe), it tattooed her back, it ignored the ponytail, and in its epoch samples the freckles and tattoos nearly vanish. Epoch 10 duplicated her.

The jump from 3 to 6 is two tags. Skin tone and freckles never stuck to the trigger, because WAI pulls toward pale skin with no freckles. Put freckles, tan skin in the prompt and those two scores go from 4.5 to 6.4.

A note on the judge: it was calibrated on her reference art, which isn't drawn like WAI, so it docks WAI's line a little. My eye made the final call.

The 8-of-10 bar still didn't close. What's missing is tattoo: sleeves at 5.4, the shoulder skull at 4.3.

Train Further didn't help

More steps felt like the obvious fix. Train Further on epoch 10, +1,000 steps, 400 Buzz. Same prompt, same tags, weight 1.0: 4/10, average 6.37, against 6.40 before. The average barely moved.

The shoulder skull went up a little, 4.3 to 5.1. Sleeves stayed at 5.4. And the skull print from her band tee started leaking onto shirts that never asked for it, the café scene included. More steps on the same images seem to learn her clothes before her tattoos. Next test: forearm and shoulder crops in close-up. That version stays a private draft, and V1 is epoch 10 of the first run.

The tan line fix

tan skin brought her color back and brought swimsuit tan lines with it. All five seeds I tried had them, and tanlines in the negative alone didn't stop it. The fix is to swap tan skin for dark skin and keep tanlines in the negative. I checked that by eye on the gallery seeds. It isn't in the scored table. WAI adds the tan lines on its own.

What I'd run

• ariamaiven, plus ariamaiven outfit for her look

• freckles, tan skin, or dark skin with tanlines in the negative

• Weight 1.0. Drop to 0.8 if a skull print shows up on clothes you didn't ask for

• WAI-Illustrious, Euler a, 28 steps, CFG 6, 832x1216

• Close-ups hold the forearm tattoos. Wide shots simplify them.

I also checked it on Civitai's own generator, WAI v17 at its defaults (Euler a, 30 steps, CFG 7), atelier scene in her look: 4 images for 18 Buzz, all four of them her. The defects were the known ones, made-up text on the tee and the skull print at 1.0.

The back view never made the model page gallery. Across five seeds none put the skull on her left shoulder, and freckles spreads freckles all over her back.

Try it: https://civitai.red/models/3022255/aria-maiven

If you've fixed a one-sided tattoo in a character LoRA, was it crops, captions, or something I haven't tried?