当前版本为(2k_epoch-4 2k_epoch-3)
Current Version: (2k_epoch-4, 2k_epoch-3)
以下为当前2k训练的模型信息:
Current 2K Model Training Information:
样本数量:=1764张
Total samples: 1764 images
样本尺寸:2560x1440 分辨率
Resolution: 2560×1440
样本远景:~200张
Long shots: ~200
样本全景:~200张
Full-body shots: ~200
样本中景:~500张
Medium shots: ~500
样本近景:~500张
Close-up shots: ~500
样本特写:~200张
Extreme close-ups: ~200
样本收集:6天
Data collection: 6 days
样本P图:15天
Image editing: 15 days
样本标注:35天
Annotation: 35 days
训练耗时:6天(RTX5090)
Training time: 6 days (RTX 5090)
(Training Tool)训练工具:DiffSynth-Studio
(Parameters)训练参数:
height 1440
width 2560
dataset_repeat 6
learning_rate 1e-4
num_epochs 3
lora_rank 32
model_id_with_origin_paths:
Tongyi-MAI/Z-Image-Turbo:transformer/*.safetensors,
Tongyi-MAI/Z-Image-Turbo:text_encoder/*.safetensors,
Tongyi-MAI/Z-Image-Turbo:vae/diffusion_pytorch_model.safetensors
===================================
标注维度:
Annotation Format:
[固定头部与服饰描述]。[动作/姿态]。[人物朝向]。[表情]。[面部动作]。[视角]。[遮挡/道具]。[人物裁切]。[景别]。[环境/光照/天气]。
[Fixed character & outfit description]。[Action / Pose]。[Character orientation]。[Expression]。[Facial details]。[Camera angle]。[Occlusion / Props]。[Framing / Crop]。[Shot type]。[Environment / Lighting / Weather]。
标注示例:
annotation Examples:
libaoping01,黑发双丸子头,红色汉服短裙白色毛绒镶边。侧坐姿,右腿收拢,右膝弯曲,左腿伸出,左腿张开,左膝微屈,左臂展于身侧支撑,右手抬于膝盖,头部转向正面。面向左前侧。严肃表情。眉毛压低,微闭眼。左脚被裁剪。全景。山林间柔和天光。
libaoping01,黑发双丸子头,红色汉服白色毛绒镶边。双手握剑于身前。面向右前侧。平静表情,微闭眼,微张嘴。仰拍。前景遮挡腰部以下。中景。山间暖橙柔光。
libaoping01,黑发,红色汉服白色毛绒镶边。右手抬起掩口鼻。微低头。面向右前侧。眉头微低,微闭眼。头顶被裁剪,肩部以下被裁剪。近景。山间暖橙柔光。
libaoping01,黑发。面向镜头。平静表情,眉毛微压低。眉毛以上被裁剪,前景遮挡鼻子以下。特写。室内暖柔光。
===================================
使用建议:
Usage Recommendations:
权重:0.6-0.8
Weight: 0.6 – 0.8
步数:30
Steps: 30
cfg:2
CFG: 2
基础模型:Tongyi-MAI/Z-Image-Turbo
Base Model: Tongyi-MAI/Z-Image-Turbo
分辨率:2560x1440/1440x1440
Resolution: 2560×1440 / 1440×1440
提示词样本:
Prompt Examples:
libaoping01,黑发双丸子头,红色汉服短裙白色毛绒镶边。侧坐姿,右腿收拢,右膝弯曲,左腿伸出,左腿张开,左膝微屈,左臂展于身侧支撑,右手抬于膝盖,头部转向正面。面向左前侧。严肃表情。眉毛压低,微闭眼。左脚被裁剪。全景。山林间柔和天光。
libaoping01,黑发双丸子头,红色汉服短裙白色毛绒镶边。Side sitting pose, right leg tucked in with the knee bent, left leg extended outward and spread, left knee slightly bent, left arm extended to the side supporting the body, right hand raised and resting on the knee, head turned forward. Facing front-left. Serious expression. Lowered eyebrows, slightly closed eyes. Left foot cropped. Full-body shot. Soft daylight in a forest.。
本lora只在官方发布的Z-Image-Turbo 上测试过,其他版本的Z-Image我没有测试过。
This LoRA has only been tested on the official Z-Image-Turbo model. Other Z-Image versions have not been tested.
===================================
此为人生中训练的第二个lora。
This is the second LoRA I have trained.
后续计划:
Future Plans:
1、训练qwen_image版本
1、Train Qwen-Image version
2、训练flux2版本
2、Train Flux2 version
3、训练wan2.2版本
3、Train Wan 2.2 version
禁止生成NSFW 内容
NSFW content is strictly prohibited.
===================================
除当前2K模型(2k_epoch-4 2k_epoch-3)外,还有 1k_epoch-4 和 1k_epoch-3 模型,相比当前2K版本做了训练优化,效果会好一点。
In addition to the current 2K model (2k_epoch-4, 2k_epoch-3), there are also 1k_epoch-4 and 1k_epoch-3 versions. These 1K versions include further training optimizations and generally achieve slightly better results.
优化内容如下:
Optimization Details:
1、删除部分远景样本。
1、Removed some long-distance (far-shot) samples
2、远景样本中不再描述五官。
2、Facial features are no longer described in far-shot samples
3、全景中部分样本不再描述五官。
3、Some full-body samples no longer include facial feature descriptions
4、标注维度顺序改变。
4、Changed annotation dimension order
5、标注开头拆解。
5、Refactored annotation prefixes
6、分辨率改成1k(1920x1080/1440x1440)
6、Resolution changed to 1K (1920×1080 / 1440×1440)
7、dataset_repeat 6 改为dataset_repeat 5
7、dataset_repeat reduced from 6 to 5
注:但是 1k_epoch-4 和 1k_epoch-3 模型 我并没有上传,因为我不打算将它作为免费使用的模型,你可以在这里获取他们。
Note: The 1K models (1k_epoch-4 and 1k_epoch-3) have not been uploaded, as they are not intended to be free.
https://www.patreon.com/posts/li-bao-ping-ren-155261467
图片展示:
Picture display:
https://civitai.com/images/127075585
