Stable diffusion thread

[AikiBoy]

I use Midjourney (paid services) and Fooocus. Fooocus is mainly for the body poses and faces because can load LoRAs into it. Then finish off with MJ for the photographic finish. You need to be precise with your prompts. Not too much, not too little.

oo i used to pay for MJ when is first came out. on discord. but it the beginning phase, i bored out.

 

[Prismatic]

oo i used to pay for MJ when is first came out. on discord. but it the beginning phase, i bored out.

They are still mainly on Discord, but online app coming soon. I've tried DALL:E3 and Leonardo, but don't really like their styling. Somehow, their images are too sharp and colourful.

It's easy to give up when the generated images don't really match what you have in mind. You need to learn how to describe what you want quite precisely, and every programs have their own quirks. Then slowly do inpainting/outpaint bit by bit to get the final picture.

My 2 recent creations:

 

[AikiBoy]

I am going to start a new channel on animals to see how well AI animals perform.
Because Kling AI has animal motions!!!

What u use to generate animal. Got what model good for animal?

What the best general model now?

 

[doogyhatts]

What u use to generate animal. Got what model good for animal?

What the best general model now?

Use KlingAI by Kuai Shou.
You need to download the Kuai Shou app in order to access KlingAI.
https://kling.kuaishou.com/en

 

[doogyhatts]

Kuai Shou now have their own image model, except it uses Unet as opposed to DiT architecture.
In the meantime, their Kling platform is now available for limited use.

ComfyUI wrapper by Kijai for Kuaishou's Kolors.

 

[doogyhatts]

I think KlingAI's T2I is likely using their own model, Kolors?

 

[AikiBoy]

I think KlingAI's T2I is likely using their own model, Kolors?

so many AI competition out there...

 

[doogyhatts]

so many AI competition out there...

You can also use Kolors to generate free images.
Same login process as Kling, since this is from the same company, Kuaishou.

Biomecha test again. Raw output from Kolors website.
This is actually very good, just need to refine it using Photon.

 

[doogyhatts]

Better facial animations from Kling compared to Dream Machine.
Now AI-girls can smile a lot more.

 

[doogyhatts]

Testing new PixArt-Sigma 900M model for 1K resolution.
Much better prompt adherence but output has less variation across different seeds.

 

[doogyhatts]

Retro SciFi test using PixArt 900M model.

 

[doogyhatts]

Removed the negative prompts.

 

[doogyhatts]

Brain Processors.
PixArt 600M 2K model.

 

[doogyhatts]

Testing ipndm sampler at 40 steps. Disabled photon refiner.
Much clearer image than Euler at 16 steps.

 

[doogyhatts]

All the new sampling techniques and variations getting confusing.

 

[doogyhatts]

PixArt-Sigma 600M 1K model, using ipndm_v and euler_cfg_pp samplers.
Using KlingAI for animation and Foleycrafter for sound effects.

Natgeo or BBC earth style.

 

[doogyhatts]

More new stuff from Matto!

 

[doogyhatts]

Open Sora Plan reaches v.1.2.0.
https://github.com/PKU-YuanGroup/Open-Sora-Plan/blob/main/docs/Report-v1.2.0.md

 

[doogyhatts]

PixArt-Sigma 600M 1K model, using ipndm_v and euler_cfg_pp samplers.
Using KlingAI for animation and Foleycrafter for sound effects.

Natgeo or BBC earth style.

The long-form video on YT reached 9.6 hours of watch time in a few days.
So Japanese mythology has been a very good choice for content production.

 

[doogyhatts]

Stability.ai released new paper on Stable Video 4D.
https://sv4d.github.io/