Stable diffusion thread

[doogyhatts]

pixart?

Yes. But refined using Photon.

 

[doogyhatts]

ODE - ordinary differential equation.
SDE - stochastic differential equation.
Also got Runge-Kutta stuff.

 

[doogyhatts]

AutoStudio (Lenovo Research) has a comic strip generation solution better than StoryDiffusion.

MOFA-video from Tencent AI Lab combines different elements together for image-to-video generation.

This is the low vram solution for converting a realistic video to smooth anime.
It uses less control-net models compared to the more complex Diffutoon.

 

[doogyhatts]

trying dream machine... wolf transform into a man... this is what i get

https://streamable.com/nuczki

Dream Machine can now input two images and transform.

 

[doogyhatts]

 

[doogyhatts]

Tutorial on how to sign up for Kling.

Spoiler

 

[doogyhatts]

New samplers and schedulers!

 

[Prismatic]

I'm a bit slow to the game so just starting out on images first.

Trying to push the limits of how much details I can get into the image without it looking weird. Experimenting with culture-specific refererences, but quite hard cos existing models quite limited.

https://www.instagram.com/nihon_josei_zukan/

 

[x1243x]

I'm a bit slow to the game so just starting out on images first.

Trying to push the limits of how much details I can get into the image without it looking weird. Experimenting with culture-specific refererences, but quite hard cos existing models quite limited.

https://www.instagram.com/nihon_josei_zukan/

first and third not bad.. just that the hand in the first pic is off.. can inpaint it

 

[Prismatic]

first and third not bad.. just that the hand in the first pic is off.. can inpaint it

Oh thanks! I didn't actually notice it the first time! Managed to remove it in ver 2.

 

[doogyhatts]

DynamiCrafter will be updated soon with a new inference solution that allows for wider motions!
Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model

After which, we can simply extend the video length by using the last frame and regenerate a connecting video segment.
Then, we don't need to depend so much on Dream Machine.

 

[doogyhatts]

Tencent AI Lab presents MimicMotion, better than its own MusePose.
Now AI-characters can do better dancing and talking, with better looking hands (separated).
Requires 23gb vram!

Here are 2 samples from the project page.

 

[doogyhatts]

Tested DynamiCrafter CIL version.
I used 945M and disabled analytic_init. Has watermark.


Waiting for non-watermarked model to be released.

 

[AZE]

As for Lumina, have to wait until they finish researching the 16-channel VAE that SD3 uses.
In the meantime, Kijai has provided a ComfyUI wrapper for it.
Maybe worth trying out their Next-SFT model first and the Gemma-2b LLM that it uses.

https://huggingface.co/AuraDiffusion/16ch-vae
https://huggingface.co/ostris/vae-kl-f8-d16

Got a few ppl/groups training their own VAEs already.
Archerly still don't know huai they creat 16 channel VAE.

Original ish 8x8 compression, channel 3-> 4, rike e.g. 64x64x3 -> 8x8x4,
16 channel ish same 8x8 compression, channel 3-16, e.g. 64x64x3 -> 8x8x16,
Huai do they not drop pixel space compression to 4x4, while keeping the channels the same? e.g. 64x64x3 -> 16x16x4.

Anyway a 16channel ones being better than a 4 channel one ish obvious, all other things being equal, they basically decrease compression by 4x anyway.

 

[AZE]

sai is gone case already

They planning for an "updated/fixed" model with new license.

 

[AZE]

Lumina T2X uses DiT architecture.
Can do different outputs, not just image.

All the modern models, Pixart, Hunyuan, Lumina and SD3 are different variants of custom DiT models.
Other than Pixart that ish specialised on T2I, all the other 3's architecture seems to be designed to be extensible for multimodal systems.

 

[doogyhatts]

Testing biomechanical creature using PixArt-Sigma.
At the moment, Midjourney still do the best in this area.

PixArt-Sigma 1K model.

PixArt-Sigma 2K model.

 

[focus1974]

Oh thanks! I didn't actually notice it the first time! Managed to remove it in ver 2.

wa.. making money with instagram model ..

is it easy?
what you use? SDXL or just SD1.5?

 

[doogyhatts]

wa.. making money with instagram model ..

I think it is not possible to make any money off advertisements even if the account reached past 10K followers.
https://help.instagram.com/5485466918184985

 

[doogyhatts]

Testing biomechanical creature using PixArt-Sigma.
At the moment, Midjourney still do the best in this area.

PixArt-Sigma 1K model.

PixArt-Sigma 2K model.

Comparing with Lumina-Next-SFT HF demo.

And Hunyuan HF demo.

Spoiler