Krea 2 itself supports normalized coordinate positioning to some extent, 0~1 or 0~1000. Its own positioning ability is luck-dependent, sometimes easily displaying coordinates as text. But IDE4 format prompts are too complex, with a lower success rate on Krea 2.

Currently the LoRA is trained only on 350 images with simple descriptions + bbox positioning, using only Attention layers to avoid overfitting.

It's more reliable than the original model for positioning most of the time, but not always consistent.

Constrained by how the LoRA training handles auto-cropping across different resolutions, some coordinates end up misaligned.

Most images position the subject in a relatively fixed spot, like centered, making unusual sizes or coordinates harder to follow—such as placing your desired subject at the top edge of the image.

Bounding boxes also have limitations: across different images, some completely cover the whole subject like a standard standing pose, others cover most but not all like excluding an extra-long tail, and some are just local like close-ups. So bounding boxes can't really confine everything perfectly within them.

Most images highlight the subject, causing the content to be rendered larger than expected.

Due to the limitations of the equipment, training speed, bounding box alignment, and the model being trained with fp8 precision, the training resolution can only achieve relatively precise bounding boxes at 640, which may make it difficult to learn small objects.

Format:

The all objects uses the following bbox.
{
  "objects": {
    "A small lying cat": [ 474, 0, 1000, 171 ],
    "a small lying dog": [ 0, 831, 481, 1000 ],
    "a square rug": [ 0, 171, 1000, 834 ]
  }
}

Actually you don't need to follow it completely, can just write the coordinates.

Tool (a powerful caption editor that supports bounding boxes):

-> Dataset Tag Editor <-

some demo images:

offset caused by image cropping in the LORA trainer:

中文说明：

krea2本身在一定程度上支持使用归一化坐标定位，0~1或0~1000. 其自身定位能力看运气，有时容易将坐标以文字方式显示.

但是ide4格式的提示词格式过于复杂，在krea2上成功率不高。

当前lora只在350张经过普通描述+bbox定位的图像上训练，为避免过拟合只在Attention层训练.

在多数时候比原始模型定位可靠，但是并不总是可靠。

受限于lora训练时，图像在不同分辨率下的自动裁剪，这会导致一些坐标对不上。

大多数图像会将主体放在相对固定的位置，比如放在中心。这导致反常的大小或坐标难以遵循，比如在图像顶部边缘放置你想要的主体。

边界框也有局限性,在不同图像中,有的是边界框完整覆盖整体,比如标准站立姿势，有的覆盖大部分,比如不包含超长的尾巴,有的只是局部,比如特写。

也就是说边界框其实也做不到完全限制在框内。

大多数图像突出主体，这导致倾向于将内容绘制的较大。

受限于设备，训练速度，边界框对齐，模型使用fp8精度训练，训练分辨率只能在640时让边界靠比较精确，这可能导致难以学到小的物体。