Regardless of the oc, make use of concepts of dresses, shoes and bondages etc.

========================
Meloris (Prisoner) | MOSS Image Character LoRA

English

Meloris is an adult fictional prisoner character: long golden-blonde hair with a fine short temple braid, blue eyes, a slender adult build, gray prison dress with a 722 chest patch, pale gray thigh-high stockings, brown lace-up Oxford shoes and a silver steel collar.

Trigger: mlrmeloris

Recommended starting strength: 0.55, especially alongside Norma at 0.55.

Keep short hair/eye/outfit anchors in the prompt, and describe the pose and restraint connections explicitly. This adapter helps character consistency; it does not guarantee correct hands, numbers, chain topology or precise anatomy.

Prompt anchor:

mlrmeloris, an adult woman, long golden-blonde straight loose hair, a short fine temple braid, clearly blue eyes, slender adult proportions, gray short-sleeved collared prison dress with a small white 722 chest patch, pale gray thigh-high stockings, brown flat lace-up Oxford shoes, a rigid silver steel collar.

Known limitations: clothing and restraint details may be bound to identity; hairstyle length and eye color can drift. In two-character or crowd scenes, uniform badges, hair colors and other features can leak between subjects. The 722 patch/mark and restraint geometry still require inspection. Strongly exaggerated facial expressions remain difficult in the tested setup.

中文

Meloris 是一名成年虚构囚犯角色：金色长发与太阳穴短细编发、蓝眼、纤细成年身材；灰色 722 标牌囚服裙、浅灰过膝袜、棕色平底系带鞋和银色钢项圈。

触发词：mlrmeloris。建议从 0.55 开始，尤其是与 Norma 同框、两支角色 LoRA 各 0.55 的组合。建议保留简短发型、眼色和服装锚点，另外明确姿势与拘束连接关系。

已知问题：角色与服装/拘束存在绑定倾向，长发长度与眼色可能漂移；多人时会串警徽、发色和服装特征。722 字样、手指、链条连接与夸张表情仍需逐图检查。

训练信息 / Training

MOSS Image Base INT8; KohakuMossTrainer; 48 selected synthetic Anima-generated images per character; natural-language captions; rank 32 / alpha 32; 1200 steps; learning rate 5e-5; AdamW; bf16 adapter training over a frozen INT8 base; native aspect ratio at approximately 1024×1024 pixel area; maximum one image per training batch; no horizontal flip. The text encoder and base weights were frozen. DMD and BunnySlop were not merged into these character adapters.

每个角色使用 48 张筛选过的 Anima 合成图，以自然语言重新标注。训练底模为 MOSS Image Base INT8，rank 32 / alpha 32，1200 步，学习率 5e-5，AdamW；保留原生宽高比、约 1024×1024 等效面积，单批最多一图，不做水平翻转。冻结底模与文本编码器，仅训练角色适配器；未将 DMD 或 BunnySlop 合并到权重中。

Tested inference / 已验证的出图配置

MOSS Image SFT INT8 + qwen3.5_9b_int8_convrot + flux2-vae. Character LoRA: 0.55 each in the two-character setup; BunnySlop_MOSS: 0.5625; MOSS_SFT_8step_DMD_latest: 0.7. 8 steps, CFG 1, Euler / Simple, denoise 1. SkipLayerGuidanceDiTSimple: single_layers=22, double_layers empty, start=0, end=1. Typical duo resolution: 640×1920; street crowd tests: 1024×1024. No artist tags, regional masks, reference-image input or ControlNet in these tests.

当前主要验证于 SFT INT8，而不是同一 Base 训练底模。双角色同框时推荐各 0.55；这不是对所有单人场景的强度最优结论。示例图若搭配 BunnySlop、DMD 或另一支角色 LoRA，会单独注明，不应把整张图的画风都归因于本角色 LoRA。

Credits / 致谢

MOSS Image: https://civarchive.com/models/3005406/moss-image

KohakuMossTrainer: https://github.com/KohakuBlueleaf/KohakuMossTrainer

BunnySlop (external style adapter used in the styled examples, not bundled): https://civarchive.com/models/2613391/bunnyslop-ai-style-moss-or-anima-or-illustrious?modelVersionId=3412969

File: meloris_moss_v1_base.safetensors

Size: 198237768 bytes

SHA256: dee57f501b61ef237df2ba2e493f7d2ae9126fb05fb93a1d35ac7ed61f63c27e