这个模型是在2509模型刚出来的时候使用大批量的姿势骨骼图来进行大批量训练的(单卡4090 48G训练时长约17-25天左右），在当时的kontext的控制并不是那么完美的时候，我急需一个既不改变背景又不改变模特包括模特身上的服装，只改变模特身上的姿势，并且这个姿势是可以被我控制的。

因此，这个lora应运而生，并且在后续的包括但不限于qwen2511跟klein出来之后我继续进行了对应的训练，反过头来发现还是以2509训练出来的姿势是最稳定，最符合我训练之初的理念。

并且这个模型训练出来之后有两个更令我惊喜的地方：

1：这个模型加了8步加速lora之后，依旧可以稳定的控制，对于画面的损失并不是想象中的那么的严重；

2：这个模型不局限于以骨骼图来进行控制，以文本也可以进行对应的姿势控制，更增添了这个lora的可玩性与趣味性！毕竟，我们最初的快乐不正是抽卡时那种未知的刺激感吗？

3：（补充条款）text-image可以用batch进行输入来达到批量出图的效果。

当然，可能这个模型还有更多的玩法，但本质上是在保持背景与服装的一致性的基础上来对模特的姿势进行指定或者随机性的改变。

这个模型的工作流以及模型我已经放到了：https://studio.aigate.cc/images/1094631428065984512?release=v0.0.2
这个镜像里面，免费使用，模型我也已经上传到里面了，工作流也在里面。有需要的可以去学习、使用。（48G起用，FP8实测显存占用约34G左右。）

参数：

lora强度：推荐0.8-1

cfg：2.5

step：20

采样器：er_sde

调度器：beta57

触发词：Based on this skeleton diagram, adjust the model’s pose to match the corresponding posture while keeping the original background and details unchanged.

工作流的搭建使用：图像推荐缩放到1024，当然1536也可以，因为这是qwen，并且你要确保你要有足够的显存：1024的显存占用大约35~36左右，而1536的显存占用会到50以上。推荐openpose或者DW骨骼图进行控制，用文本控制的方式为：Change the pose to xxxxxxxxxxx, while maintaining the same subject and composition.

模型的描述到这里也大抵结束了，感谢大家的支持，如果有其他的需求，可以联系我，当然，这不一定免费，谁知道呢？

QQ:1154625224

微信：z1154615114

This model was trained with a large number of pose skeleton images when the 2509 model had just been released. The training took about 17–25 days on a single RTX 4090 (48GB). At that time, the control capability of Kontext was not yet perfect, and I urgently needed something that could change only the model’s pose while keeping the background, the model, and the clothing completely unchanged, with the pose being fully controllable.

Thus, this LoRA was created. Later, when models such as qwen2511 and klein came out, I continued with corresponding training. Looking back, I found that the poses trained based on 2509 are still the most stable and best aligned with my original training philosophy.

After training this model, two more things pleasantly surprised me:

1. After adding the 8-step acceleration LoRA, it can still maintain stable control, and the loss in image quality is far less severe than expected.

2. This model is not limited to skeleton-based control. It can also control poses through text, which greatly increases its playability and fun! After all, isn’t our original joy the sense of uncertainty when pulling a gacha?

3. (Additional note) For text-to-image, batch input can be used to achieve batch image generation.

Of course, there may be more ways to use this model, but its core purpose is to specify or randomly change the model’s pose while keeping the background and clothing consistent.

The workflow and model have already been uploaded to: https://studio.aigate.cc/images/1094631428065984512?release=v0.0.2
They are available for free use within this image. The model has been uploaded, and the workflow is also included. Anyone interested can go and learn from it or use it. (Requires at least 48GB of VRAM; in FP8, the tested VRAM usage is around 34GB.)

Parameters:

• LoRA strength: recommended 0.8–1

• CFG: 2.5

• Steps: 20

• Sampler: er_sde

• Scheduler: beta57

• tigger：Based on this skeleton diagram, adjust the model’s pose to match the corresponding posture while keeping the original background and details unchanged.

Workflow setup:

• Recommended to scale images to 1024 (1536 is also possible since this is Qwen), but make sure you have enough VRAM.

• VRAM usage:

  • 1024 → about 35–36GB

  • 1536 → over 50GB

• OpenPose or DW pose is recommended for control.

• Text-based control format:

  Change the pose to xxxxxxxxxxx, while maintaining the same subject and composition.

That’s basically the end of the model description. Thank you all for your support. If you have other needs, feel free to contact me — though it may not be free, who knows?

QQ: 1154625224
WeChat: z1154615114