Reposted from Huggingface ————————————

Introduction

Creating poses for Qwen Image Edit (2511) or its variants can be challenging. Even though the model has OpenPose integrated, the results can still look off—such as incorrect depth or poses that don't quite match the character, resulting in unexpected distortions. Additionally, manually creating poses can take too much time, especially if you opt to use Blender OpenPose rigging. So, what if you could skip OpenPose entirely and simply write a straightforward (though somewhat long) prompt to replicate the pose from any image on your computer?

That is what AnyPose LoRAs aim to achieve. Designed specifically for the new Qwen Image Edit 2511 Lightning LoRA to enable fast inference, it allows any image to follow a target pose using just a reference image as a pose guide. No ControlNet required.

Quick Start

Trained on Qwen Edit 2511 using a 4-step lightning LoRA. Set the weights for both AnyPose LoRAs (base and helper) to a strength of 0.7. Upload two images: one featuring the initial input character, and the other containing the pose you want to replicate onto the first image. Use the following prompt: "Make the person in image 1 strike the exact same pose as the person in image 2. Do not change the style and background of the person in image 1, so please refrain from doing so. The new pose should be pixel-accurate to the pose we want to replicate. The positions of the arms, head, and legs should match the pose to be replicated. Adjust the FOV and angle to perfectly match image 2. The head tilt and eye gaze pose should align with the person in image 2." If additional context is needed (such as clothing the model hasn't seen before), add it to the end of the prompt. Is the background inconsistent with the initial input image? Add the following at the end: "Remove the background of image 2 and replace it with the background of image 1."

Examples

AnyPose works well for simple poses, such as T or A poses, or more complex poses, such as various yoga poses:

Important Notes

Before using this tool, there are a few caveats to keep in mind. Creating the LoRA dataset sounds straightforward, right? Simply pose two characters in the exact same pose (I used Blender to pose dense 3D characters into various poses Qwen Edit couldn't understand on its own, like handstands, bridge poses, bending over, etc.); so that step was easy to complete.

However, deciding what the final processed image should look like turned out to be extremely difficult. Should the final output be the original scene with the character in a new pose? Or should the character be "replaced" into the scene of the second image? The first approach sounds the most intuitive, but it comes with drawbacks, such as the background being filled in incorrectly. Meanwhile, the second approach becomes impractical if you simply want to change the person's pose within the current scene. It really is quite unintuitive. Therefore, I tried both approaches.

In retrospect, I don't think this was the best choice, because in some cases, the character replaces the person in the second image and inherits its background—an unintended consequence; or elements from the previous scene clip into the current scene. However, much of the time, it does manage to maintain the same scene. From a consistency standpoint, this doesn't sound great and can certainly be frustrating, as it ought to "just work." The reality is that it's simply very subtle. Nevertheless, its existence brings many advantages, such as always being able to fill in missing parts.

For example, if you want to adjust the pose of a character whose full body isn't shown (such as the lower half), AnyPose will automatically blend from the original image to fill in the "unknown" areas:

To fix this issue, you can append a prompt to the end of the trigger prompt to add or remove elements in the output. For example, for unknown areas, you can specify what the character is wearing, like "the woman in image 1 is wearing white leggings":

However, even though this workaround works, please note that the best way to use AnyPose is to present the full body of the initial character in the initial image (along with a complete environment like the floor), so the generated pose remains consistent without requiring Qwen to guess.

Another point to note, which you might be thinking right now: the environment changed. In this example, I wanted the pose performed within the original background of the initial image; I didn't want the character swapped into another image's background. So how do you restore the original background? It's actually quite simple. We just need to prompt the model to swap the background of the second image with the background of the first image. For example, we can append "Remove the background of image 2 and replace it with the background of image 1." to restore the original background:

You can append a prompt at the end to resolve almost any issue that arises. Background didn't remain unchanged? Add it in the prompt. Missing a floor or the character is floating? Add it in the prompt. Character holding an object they shouldn't be? Remove it via the prompt. Character has unwanted features, like long hair, a different hair color, or similar? Remove them via the prompt.