Toggle to show / hide eyes and nose nodes

#4
by adunato - opened

When characters are seen from behind the openpose render should not display eyes and nose (only ears) within the face information. Ideally when rotating the character this would happen automatically, but it would probably best and easier to develop for the user to show/hide these nodes manually via dedicated checkbox.

How about using Q + click? You can trim from there. Note that the nose cannot be trimmed

Thanks for the suggestion, unfortunately it doesn't work as for rear position to work it should only show ears and hide nose and eyes. If not possible, then this doesn't work for control net rear position.

I think that:

1.Generating rear-facing poses is difficult even without using ControlNet.
2.ControlNet seems to use the left and right bones to estimate the orientation of the body.

Therefore, I don't think that function is very valuable for this situation.

How about trying to manually remove the unwanted nodes first? If it works well, please send me the image and pose data. That way, I can better understand what you need specifically.

Thanks, I agree with #1 however, after spending some time trying to generate rear-facing poses, having a correct (hidden nose / eyes nodes) seems to be the most effective way and in my experience it's quite effective. I don't really share the same experience regarding your 2nd point, in fact in mine left-right bones does not affect the output pose at all, although I'm sure it does have some effect in certain situations.

Not sure what you meant by "trying to manually remove the unwanted nodes first", do you mean via photoshop or similar? Happy to do that if that's what you are looking for alongside some examples of how controlnet interprets openpose inputs.

Sure, that's correct. If that's what you're looking for, you can create an ideal pose image using tools like Photoshop, and use that to generate images. If it works well, please provide both the pose image and generated images to me. That way, I can understand exactly what you're looking for without any confusion.

Here's the modified mask (eyes and nose removed) and some outputs using a variety of models (vanilla, protogen) and pruned and non-pruned openpose controlnet models, with "a man" as the only prompt. It works very reliably in my experience, it may struggle if you are using models that are heavily trained on front portraits, but that's just how stable diffusion works. Again, I think it's worth adding, but personally I work on photoshop anyway so it doesn't cost me much to edit it manually.

man.png

00592-3233935825-a man.png

00593-2803226273-a man.png

00594-2055389426-a man.png

Since I had an idea of how to do it, I'll consider it and see if it works well.
By the way, what happens if you don't erase the eyes on the same model?

Here you go, with the character turned 180d but eyes and nose visible result is front pose for all of them. I did manage to turn the model but including "view from behind", but that misses the point of this feature which is to achieve reliable back poses without relying on prompt, which is by nature limiting (multi-characters confusion, unreliable/variable outcome, conflict with other prompt elements, etc).

00595-1367669682-a man.png
man_2.png
00596-2782643790-a man.png

Thank you for the information.

Added H+click. Toggles on a node. A hollow circle will be displayed, but it will become completely blank upon saving.

Great! I'll test it tomorrow.

Fantastic, it works great and it's a very elegant way of implementing.

jonigata changed discussion status to closed

Sign up or log in to comment