Stonelax again,

I made a quick Flux workflow of the long waited open-pose and tile ControlNet modules. (Canny, depth are also included.) The backbone of this workflow is the newly launched ControlNet Union Pro by InstantX.

I quickly tested it out, anad cleaned up a standard workflow (kinda sucks that a standard workflow wasn't included in huggingface or the loader github repo) so ya'll can have a try for yourselves. Some quick impressions:

1) ControlNet Union Pro seems to take more computing power than Xlab's ControlNet, so try and keep image size small.

2) Openpose works, but it seems hard to change the style and subject of the prompt, even with the help of img2img. For example, I inputted a CR7 siu pose and inputted "a robot" in prompt, the output image remained a male soccer player. I had to lower the strength to ~0.2 and finally got a robot, but the pose was slightly off.

Comparison below:

Left - strength ~0.2, pose is slightly off                             Right - strength ~0.5, pose is accurate but no robot

3）The strength of image composition control seems to be slightly better than that of Xlab, but to be honest Xlab's Canny and Depth are quite usable already.

Anyway, having openpose and tile support is a win regardless! I will try to see if speed and style transfer can be optimized tomorrow.

Please let me know any of you make progress on speeding it up & style transfer too!

Cheers