Born on Meridian Station — a decommissioned orbital rig drifting at the L4 point between two warring colony worlds. The station was supposed to be temporary. It wasn't. Three generations of stranded laborers turned it into a permanent scrap heap of welded corridors, leaky hydroponics, and desperate people.

Nekara Vic is a character LoRA — a hardened bounty hunter with a consistent, detailed design that carries across any scene or outfit you place her in. Dark skin, a prominent scar across the right eye, thick dreadlocks (cultural — Meridian Station tradition), metal breastplate, high-cut leotard, thigh plates, and a red cape. She's tall, athletic, and moves like someone who precalculates escape routes.

She is not a generic sci-fi warrior. She is this woman — specific, consistent, recognizable. Place her in armor on a desert cliff and she owns the horizon. Place her in an evening gown at a high-society gala and she looks out of place in exactly the right way. Place her in a station bar in a tank top and cargo pants and she blends in — except for the scar.

The face is the anchor. The LoRA was trained on multiple outfits — armor, EVA suit, casual clothes, ritual garments — and the face remains Nekara across all of them. Dress her in whatever the scene demands. She'll still read as herself.

Available on five base models, each tuned to its prompt style:

• Illustrious — tag-based, anime-influenced rendering. Sharp linework, clean armor details, the scar reads especially well.

• Klein 9B — natural language, cinematic photorealism. Grain, atmosphere, skin texture, environmental lighting. Face consistency across outfit changes is this version's strength.

• Flux.1 Dev — natural language, maximum photorealistic fidelity. Detailed skin texture, accurate lighting, strong face lock. Larger LoRA file (r32/a32) but strongest character commitment.

• Z-Image — natural language, balanced photorealism. Good face consistency, softer rendering style than Flux. Step-1250 sweet spot selected for sharpest faces. Compact file size.

• Chroma — natural language, compact photorealism. Required the flowmatch noise scheduler. Step-1500 sweet spot. Slightly less detailed than Flux or Klein but good face consistency in a small package.

Recommended Settings

• Trigger: nekara_vic

• Strength: 0.6–1.0

Weight Guide

• 0.4–0.6 — Subtle influence — face suggestion, mild outfit echo

• 0.6–0.8 — Sweet spot for most scenes — strong face, flexible outfits

• 0.8–1.0 — Full character lock — best for armor scenes and direct portraits

Prompt Tips

• Resolution: 832×1216 (portrait) or 1216×832 (landscape)

• Illustrious uses booru tags — keep it focused. A few precise tags beat a wall of descriptors

• Klein/Flux/Z-Image/Chroma use natural language — describe the scene and what she's wearing, the LoRA will keep her face

• Outfit flexibility is a feature — the character was trained on armor, EVA, casual, and ritual outfits. The face is consistent across all of them. Push the LoRA by dressing her in whatever the scene demands

• The scar is her signature — it reads at all weights above 0.6, and is most prominent at 0.8+

Example Prompts

Armor Portrait (Illustrious):

nekara_vic, 1girl, solo, dark skin, scar across eye, dreadlocks, metal breastplate, high-cut leotard, thigh plates, red cape, standing in spaceship corridor, arms crossed, dramatic lighting, masterpiece, best quality

Armor Portrait (Klein 9B / Flux / Z-Image / Chroma):

nekara_vic, a dark-skinned bounty hunter woman with a scar across her eye and dreadlocks, wearing a metal breastplate, high-cut leotard, thigh plates, and a flowing red cape, standing in a spaceship corridor, arms crossed, dramatic lighting, photorealistic

Desert Canyon:

nekara_vic, a dark-skinned bounty hunter woman with a scar across her eye and dreadlocks, wearing a metal breastplate, high-cut leotard, thigh plates, and a flowing red cape, standing on a desert canyon cliff, golden hour sunlight, photorealistic

Evening Gown Outfit Swap:

nekara_vic, a dark-skinned woman with a scar across her eye and dreadlocks, wearing a deep red velvet evening gown, standing at a high-society gala, looking out of place among the wealthy, photorealistic

Casual Bar Scene (Illustrious):

nekara_vic, 1girl, solo, dark skin, scar across eye, dreadlocks, tank top, cargo pants, sitting in spaceport bar, neon lights, cyberpunk, masterpiece, best quality

Multi-Base Comparison

• Illustrious v1 — r16/a16, 3000 steps, booru tags. Best for stylized illustration, sharp linework.

• Klein 9B v2 — r32/a16, 3500 steps (best @ 1250), natural language. Best for cinematic realism.

• Flux v1 — r32/a32, 4500 steps, natural language. Best for maximum realism. Largest LoRA file (328 MB).

• Z-Image v1 — r16/a16, 3500 steps (best @ 1250), natural language. Best for balanced realism in a compact file (82 MB).

• Chroma v2 — r16/a16, 3500 steps (best @ 1500), natural language. Best for compact realism. Requires flowmatch scheduler (107 MB).

Training

All five models were trained on the same 57-image dataset using Prodigy+Schedule-Free (PPSF) at d0=1e-5 with EMA (decay 0.99). The key finding: PPSF finds a learning-rate plateau early, and the best character checkpoints come from before the LR climbs past that plateau — overtraining gradually softens face distinctiveness while sharpening outfit details.

• Klein 9B v2 @ 1250 — Steps 1000 and 1250 were close, but 1250 had slightly stronger individual features. By 1500–2000, the face was softening as PPSF pushed LR past the ~0.000466 plateau.

• Z-Image v1 @ 1250 — Same pattern as Klein. Step 1250 had the sharpest face features and best scar visibility.

• Chroma v2 @ 1500 — Needed more training to reach its sweet spot. Step 1000 was undertrained, 1250 was good, 1500 had the sharpest features. Requires flowmatch noise scheduler (not ddpm).

• Flux v1 final — Higher rank (r32/a32) meant slower convergence. The final checkpoint at 4500 steps was the strongest.

• Illustrious v1 final — Strong throughout the 3000-step run. No early-stop needed.

Full gallery + prompts: thalis.dev/models/nekara-vic

Browse all models: thalis.dev/models