Stable diffusion thread

[doogyhatts]

vram requirements keep going up

And so are the model sizes. Now must clear some disk space once in a while.

 

[doogyhatts]

Get ready for the next big leap in AI video generation!
https://openai.com/sora

 

[doogyhatts]

CNY demo from OpenAI Sora page.

 

[doogyhatts]

At the moment, it is not realistic to believe Stability.Ai is going to catch up with OpenAI.
So the limitations are now obvious.
We can only do so much with SD, SVD, AnimateDiff, etc.

Spoiler

 

[order66]

very long never play with AI liao.
was watching this clip

Hope that AI can generate text to JAV videos by the end of the year!

 

[doogyhatts]

very long never play with AI liao.
was watching this clip

Hope that AI can generate text to JAV videos by the end of the year!

Currently, I am testing an updated video-to-video solution.
But it requires existing motion capture for openpose and depth images.
It will take at least two days to make a 30 secs video of which the dance animation sequence occurs for only 5 secs.
So I have to repeat the video generation for 5x2 times with different fashion styles (2 clips per slide).

Btw, I found something else interesting for image-only videos.
The YT channel known as AI_collection_girl is approved for monetisation and uploads four slideshow videos everyday.
In that manner, the YT algorithm did take note of the consistency and actively promotes the channel's content.
And this channel started from Oct last year.

 

[doogyhatts]

 

[zheng]

 

[doogyhatts]

Just found out that AnimateLCM also has an adaptation to use SVD-xt.
But the inference code has not been released yet.
https://huggingface.co/wangfuyun/AnimateLCM-SVD-xt

And one for portraits as well.
https://huggingface.co/wangfuyun/AnimateLCM-I2V

So I have been using the text-to-video one. It is fast to generate 120 frames and has a high success rate.
It is now possible to make multiple 30 second YT videos within a day.

 

[doogyhatts]

Bytedance releases SDXL-Lightning.
ComfyUI supported!
https://huggingface.co/ByteDance/SDXL-Lightning

 

[doogyhatts]

Stability.ai announces Stable Diffusion 3!
https://stability.ai/news/stable-diffusion-3

 

[zheng]

Stability.ai announces Stable Diffusion 3!
https://stability.ai/news/stable-diffusion-3

stable cascade n sd3 are different model/product right?

 

[doogyhatts]

stable cascade n sd3 are different model/product right?

Yes, they are different.

SD3 can handle text.

 

[doogyhatts]

After further testing, I found out that AnimateDiff with LCM LoRA & new sd15_t2v_beta model can generate 40 frames video in less than one minute on my local machine.
Settings: lcm sampler, sgm_uniform for scheduler and lcm >> sqrt_linear for beta_schedule.

 

[doogyhatts]

 

[doogyhatts]

 

[zheng]

released?

 

[doogyhatts]

released?

Not yet. Got wait list.
But if you paid for the membership, I think you get to access their unreleased draft versions. Not sure if that is the case either.

 

[doogyhatts]

 

[doogyhatts]