这是面向 Krea 2 的 DiT 主干训练的角色 LoRA,单文件
yinyuexjm-raw_c1-st4000.safetensors占 223.81 MiB,全量 BF16,rank 32,可训练参数 117.31M,由 ZIYUE5516 经 ModelScope × MuseAI 于 2026 年 10 月 24 日产出,标称底模A1375569/Krea-Raw,训练 4000 步、学习率 1e-4,分辨率按约 1MP 分桶(max_pixels 1048576)。它的注入面铺得很满:28 个transformer_blocks的注意力与 FFN 全部通道(含to_gate闸门)加text_fusion的 refiner、layerwise 两组,再加密img_in、txt_in、time_embed、time_mod_proj、final_layer等 8 个全局模块,合计 264 个目标、528 张量,意味着它能改写的不仅是五官,还有体态、发丝走向和服饰结构,代价是权重做到两百多兆。形象取自《凡人修仙传》中银月化形后的三维动画造型——银白长发、兽耳、异色瞳、白底金纹长袍束朱红绶带、赤足配金饰,用触发词yinyuexjm加一组外貌辅助标签即可稳定唤出,强度落在 1.0–1.3 区间,1.2 附近才把建模质感压满;由于头部未记录lora_alpha与标签频率表,等效缩放和训练集构成只能靠换底模时实测校准,这也是它更适合绑定 Krea 2 系融合底模、而非直接套进任意 DiT 的原因。规格之外,它是一份干净的纯 adapter——没有底模切片、没有量化版本、没有内嵌触发词,223.81 MiB 的每一字节都是对 6144 宽主干的增量修正。This is a character LoRA trained against the Krea 2 DiT trunk, shipped as a single 223.81 MiB file named
yinyuexjm-raw_c1-st4000.safetensorsin full BF16, rank 32, carrying 117.31M trainable parameters. ZIYUE5516 produced it through ModelScope × MuseAI on 24 October 2026 against the declared baseA1375569/Krea-Raw, at 4000 steps and a learning rate of 1e-4, with resolution bucketed around roughly one megapixel (max_pixels 1048576). Its injection surface is broad rather than surgical: every attention and FFN channel across 28transformer_blocks(gate projections included), both the refiner and layerwise stacks insidetext_fusion, plus eight global modules —img_in,txt_in,time_embed,time_mod_proj,final_layeramong them — for 264 targets and 528 tensors in total. That breadth buys control over posture, hair flow and garment structure, not just facial identity, and it is what makes the artifact run past two hundred megabytes. The subject is Yin Yue in her 3D-animation form from A Record of a Mortal's Journey to Immortality: silver-white long hair, beast ears, heterochromia, a white gold-traced robe bound with a vermilion sash, bare feet with gold ornaments. The tokenyinyuexjmplus a short appearance tag string summons her reliably, with strength effective between 1.0 and 1.3 and around 1.2 required before the modeled-render texture fully lands. Because the header records neitherlora_alphanor a tag-frequency table, equivalent scaling and dataset makeup can only be calibrated empirically when swapping bases — which is why it belongs on Krea 2 family merges rather than being dropped into an arbitrary DiT. Beyond the spec sheet it is a clean pure adapter: no base-model slices, no quantized build, no embedded trigger words, every one of its 223.81 MiB an incremental correction layered onto a 6144-wide trunk.


