Hard to generate full-body shot

[lanson07]

Amazing work！
We have consistently observed that the image generation model exhibits a strong tendency to produce headshots for single-person scenes. Despite our repeated efforts to specify prompts such as "full-body shot," "person from head to toe," or descriptive actions like "standing in a room," the generated results remain predominantly tightly-cropped portraits.
What is the specific solution or required prompt syntax to correct this and consistently generate full-body images?

[Doby-Xu]

Thanks a lot for your attention! ❤

WithAnyone has no problem generating full-body shots under proper conditions. Have you tried manually defining the bboxes for the model? The bounding box determines where and how large the face region will be.

In our Gradio demo, you can specify bboxes in two ways:

1. Upload an image with the desired number and locations of faces.
2. Manually input the bbox coordinates in the format: x1, y1, x2, y2.

You can try uploading a full-body image and using its face location for generation! Or try locating the bbox at the top of a 9:16 portrait, like top 1/8 of an image with 512 width and 1024 height.

If no bboxes are provided, the Gradio demo will randomly select one from a predefined set — and note that none of these include full-body settings.

If you have already provided bboxes but still cannot generate full-body images, please continue the discussion by sharing your generation hyperparameters so we can help reproduce and diagnose the issue.