Stable diffusion thread

[doogyhatts]

Testing a 4 second animation using Haiper.ai.
It generated three foot steps and the feet looks much better because it actually altered the feet closer to that of a panda.

Spoiler

 

[doogyhatts]

I think Haiper.ai will not be free eventually.
So back to more testing using DynamiCrafter.
This time, the two foot steps animation comes with a change in facing direction to its diagonal right-side.

Spoiler

 

[doogyhatts]

Spent some time trying out Ideogram.
Output quality for non-humans is a lot closer to Midjourney than Leonardo.
And comes with 25 free prompts per day.

 

[zheng]

Spent some time trying out Ideogram.
Output quality for non-humans is a lot closer to Midjourney than Leonardo.
And comes with 25 free prompts per day.

yah think is very good for poster, logo, text base stuff.

can do copyright characters too

actually very good

 

[doogyhatts]

I used Ideogram to generate a draft scene and character pose before redoing it in Fooocus.
This one is a draft scene.

 

[zheng]

I used Ideogram to generate a draft scene and character pose before redoing it in Fooocus.
This one is a draft scene.

need to inpaint her face

 

[doogyhatts]

I couldn't get it to spell Orange Marmalade.

 

[Loratadine10mg]

The India one is very nice

What India?

 

[doogyhatts]

Motion director tutorial. Need 14 gb vram.

 

[zheng]

Motion director tutorial. Need 14 gb vram.

still very garbage lah not worth the time to do training

 

[doogyhatts]

still very garbage lah not worth the time to do training

The output from the github page looks good though.
Not sure how come their output is better.

 

[doogyhatts]

2 second animation test using Haiper.ai, for a Japanese yokai, the Kappa.

 

[doogyhatts]

 

[doogyhatts]

This model can generate images with cinematic still quality.

 

[doogyhatts]

Ideogram cannot use image prompt at free tier, so I discovered that Haiper can generate ukiyo-e art style using its text-to-video.
Then I input it as image prompt in Fooocus and got a high detail rabbit art in painting style.

 

[doogyhatts]

T-Gate increases performance of generation.
https://github.com/HaozheLiu-ST/T-GATE

 

[doogyhatts]

Tencent AI Lab releases MusePose weights and inference code.
Need 22gb vram!
https://github.com/TMElyralab/MusePose

 

[doogyhatts]

Their MuseV and MuseTalk projects also look cool.
MuseV is for I2V and V2V. I2V can generate video with 10 seconds. It was trained on 512x320 resolution.
MuseTalk is for lip sync, used together with MuseV.

 

[doogyhatts]

PixArt Sigma requires 18gb vram!
This architecture is closest to SD3.

 

[doogyhatts]

Testing the online demo for PixArt-Sigma.
It uses the XL-1024 model, but their model download page has the XL-2K model as well.

I couldn't get wide shot angle to work in the online demo.