Doll Style / BJD Photography LoRA | Anima ver.
这是一个基于 Anima base v1.0 训练的 LoRA 模型，专注于生成 BJD (Ball-jointed doll) 娃娃摄影风格的图像。
This is a LoRA model trained on Anima base v1.0, specializing in generating images in a BJD (Ball-jointed doll) photography style.
💡 如何使用 | How to use
无触发词，直接使用 | No trigger word required, use directly.
权重:
1.0| Weight:1.0
1. ⚙️ 参数设置 | Settings (CFG, Steps)
CFG和Steps等参数可以灵活调整，但我推荐从 CFG 4~5、Steps 30 开始尝试。具体设置请参考示例图片的元数据（Metadata），图像中包含了 ComfyUI 工作流信息。Parameters such as
CFGandStepscan vary, but I recommend starting with CFG 4~5 and 30 steps. For specific settings, please refer to the metadata of the sample images, which contain ComfyUI workflow information.
2. ✨ 高清修复 | Hires Fix
开启 Hires Fix后，可以更好还原 BJD 的皮肤、发丝纹理和眼部细节。
With Hires Fix enabled, it can better restore the skin texture, hair strand details, and eye details.
Important reminder: Do not use for commercial purposes or publish on other public platforms
重要提示：请勿商用并发表到其他公开平台
Description
Base: Anima base v1.0
Training parameters:
resolution: 1024
reg_data_dir: null
reg_caption: null
reg_weight: 1.0
shuffle_caption: true
keep_tokens: 0
flip_augment: true
tag_dropout: 0.0
prefer_json: true
cache_latents: true
lora_type: lora
lora_rank: 32
lora_alpha: 32.0
lokr_factor: 8
lora_dora: true
lora_rs: false
lora_dropout: 0.0
lora_rank_dropout: 0.0
lora_module_dropout: 0.0
epochs: 15
max_steps: 0
batch_size: 2
grad_accum: 1
learning_rate: 1.0
lr_scheduler: none
lr_scheduler_t0: 500
lr_scheduler_t_mult: 2.0
lr_scheduler_eta_min: 1.0e-06
optimizer_type: prodigy_plus_schedulefree
prodigy_d_coef: 1.0
prodigy_safeguard_warmup: true
ppsf_d_coef: 1.0
ppsf_prodigy_steps: 0
ppsf_beta1: 0.9
ppsf_beta2: 0.99
ppsf_split_groups: true
ppsf_split_groups_mean: false
ppsf_use_speed: false
ppsf_fused_back_pass: false
ppsf_use_stableadamw: true
weight_decay: 0.0
kv_trim: false
noise_offset: 0.0
pyramid_noise_iters: 0
pyramid_noise_discount: 0.35
timestep_sampling: logit_normal
timestep_shift: 3.0
loss_weighting: min_snr
min_snr_gamma: 5.0
weight_cap_ratio: 0.0
grad_clip_max_norm: 0.0
mixed_precision: bf16
grad_checkpoint: true
attention_backend: none
num_workers: 0