Stable diffusion thread

[doogyhatts]

I understood a bit more on how those dancing humans and animals might be created using those apps.
I think they use a lot of segmentation and it is much faster to change the different body parts and clothing using a pre-built animation.

To make the pre-built animation, I think only DragDiffusion is the correct solution. That means, manually dragging each limb of the cat or human until it fits nicely within the reference keyframe.

But DragDiffusion can only be installed on a Linux machine and requires at least 14gb vram. So a lot of people cannot use it unless they pay for Colab or rent GPU on other cloud services. The free Colab does not work anymore as the Colab AI will attempt to crash the session.

The simpler option is to pay for the app service but will be limited to specific dance styles that the app developer has added.

Another possibility is the open source version of AnimateAnyone. But this requires 24gb vram and the solution has errors.

 

[doogyhatts]

Shirt replacement test.
HyperSD 1.5 (right-side) vs AlignYourSteps (left-side).

Spoiler

AYS somehow botched the center part.
So HyperSD 1.5 has a better output.

 

[zheng]

Shirt replacement test.
HyperSD 1.5 (right-side) vs AlignYourSteps (left-side).

Spoiler

AYS somehow botched the center part.
So HyperSD 1.5 has a better output.

cant tell whats the different

 

[doogyhatts]

cant tell whats the different

The skirt has a slight protrusion for the left-side.

 

[doogyhatts]

I realised my masking size is incorrect and had to be increased.
Here are the results for frame 28.
Left side is HyperSD1.5, while right-side is AYS.

The side effect is that all the shirts will have a big collar.
The output for HyperSD has slightly less pink and AYS one is slightly more saturated.

Spoiler

 

[doogyhatts]

 

[doogyhatts]

Trying out Fooocus again, this time on Alice character.
Not bad, at least got the Analog Film style. I added the technicolor style LoRA as well.

Now to figure out how to do face swap to be Asian looking.

 

[zheng]

Trying out Fooocus again, this time on Alice character.
Not bad, at least got the Analog Film style. I added the technicolor style LoRA as well.


Now to figure out how to do face swap to be Asian looking.

inpaint with a chinese name in prompt

 

[GhostZM]

Trying out Fooocus again, this time on Alice character.
Not bad, at least got the Analog Film style. I added the technicolor style LoRA as well.

Now to figure out how to do face swap to be Asian looking.

Looks suitable to be made into a horror film..lol


CheerS!!!

 

[doogyhatts]

Ok I managed to convert a caucasian looking face to an asian one.
I used inpainting with prompts and ChilloutMix model. Then do face fix with ADetailer.

This test is now complete. Now to read the ebook and figure out the scenes.

 

[x1243x]

Ok I managed to convert a caucasian looking face to an asian one.
I used inpainting with prompts and ChilloutMix model. Then do face fix with ADetailer.

This test is now complete. Now to read the ebook and figure out the scenes.

not going to inpaint the hands?

 

[doogyhatts]

not going to inpaint the hands?

For this test, I will ignore the hands as I intend to use it for banner and avatar only.

The actual scenes are somewhat more complex as there two girls sitting on the grass bank, reading book, etc.
And then have to figure out how to do the running animation for the white rabbit.

 

[x1243x]

For this test, I will ignore the hands as I intend to use it for banner and avatar only.

The actual scenes are somewhat more complex as there two girls sitting on the grass bank, reading book, etc.
And then have to figure out how to do the running animation for the white rabbit.

I'm playing around with svd but the quality not very good

 

[doogyhatts]

I'm playing around with svd but the quality not very good

SVD has a lot of problems with hands. Even with only upper body shots, the output can be a mess.
Have you tried DynamiCrafter? It is much better than SVD.
But I kept SVD for other purposes (eg water, fire, etc).

 

[x1243x]

SVD has a lot of problems with hands. Even with only upper body shots, the output can be a mess.
Have you tried DynamiCrafter? It is much better than SVD.
But I kept SVD for other purposes (eg water, fire, etc).

does dynamicrafter work on forgeui?

 

[doogyhatts]

does dynamicrafter work on forgeui?

I used ComfyUI for it.
There are a few wrappers. Chaojie. Kijai.
So far I have tried the Kijai one. I haven't tried the Chaojie's version.

 

[x1243x]

I used ComfyUI for it.
There are a few wrappers. Chaojie. Kijai.
So far I have tried the Kijai one. I haven't tried the Chaojie's version.

I have comfy also.. but I very sian to use it.. haha may try it again in the future

 

[doogyhatts]

I have comfy also.. but I very sian to use it.. haha may try it again in the future

It takes a long time to identify the correct setup for a specific workflow, especially for AnimateDiff ones.
For SVD and DynamiCrafter, the workflows are much shorter.

 

[doogyhatts]

Imposing Consistent Light
From the authors of Controlnet.
ComfyUI wrapper.

 

[doogyhatts]

I forgot to test noise effect and retro film filter in CapCut.