Stable diffusion thread

[zheng]

actually can do it here too
https://www.meta.ai/

 

[doogyhatts]

 

[doogyhatts]

Morphing animation is back on trend, this time using Animate LCM and output is of sufficient high quality and non-blurry.
Workflow:
https://civitai.com/models/372584?modelVersionId=426894

Might be better to classify the style as surrealistic or psychedelic morphing animation.

 

[doogyhatts]

Morphing animation is back on trend, this time using Animate LCM and output is of sufficient high quality and non-blurry.
Workflow:
https://civitai.com/models/372584?modelVersionId=426894

Might be better to classify the style as surrealistic or psychedelic morphing animation.

The workflow is so awesome!
It is very fast to generate these morphing-based animations.

Here is an example of a surrealistic animation of a face close-up. Non-upscaled.

 

[zheng]

The workflow is so awesome!
It is very fast to generate these morphing-based animations.

Here is an example of a surrealistic animation of a face close-up. Non-upscaled.

cool bery good consistency

 

[doogyhatts]

New paper from NVidia!
Align Your Steps


Looks like they fixed SVD as well with this paper.
However, after testing it again, SVD still performs worse than DynamiCrafter.

 

[doogyhatts]

Ok I managed to get some of the results with AYS + PAG.
AYS improves the quality at low steps, so that the latent upscale requires less steps.

It seems PAG can help to fix the hands but requires AYS, and is not perfect either.
I haven't tried the Smea sampler yet.

Result of PAG + AYS + Latent Upscale:

Spoiler

Removing either PAG or AYS will result in worse quality for the same seed.

 

[zheng]

Ok I managed to get some of the results with AYS + PAG.
AYS improves the quality at low steps, so that the latent upscale requires less steps.

It seems PAG can help to fix the hands but requires AYS, and is not perfect either.
I haven't tried the Smea sampler yet.

Result of PAG + AYS + Latent Upscale:

Spoiler

Removing either PAG or AYS will result in worse quality for the same seed.

hand still problematic

 

[doogyhatts]

hand still problematic

It looks better now after I try using the hand refiner.

Spoiler

 

[doogyhatts]

Hyper-SD.
https://hyper-sd.github.io/

Also have SD1.5 8-step model but I could not get a good result.
It works great on SDXL. Tried it using Fooocus.

 

[zheng]

Hyper-SD.
https://hyper-sd.github.io/

Also have SD1.5 8-step model but I could not get a good result.
It works great on SDXL. Tried it using Fooocus.

just saw this oso. havent try it

 

[Scarlet Tea]

need some help. whenever i try to access stable diffusion thru command prompt i always get this error. when i press any key the cmd prompt just disappear. im using comfyui

 

[doogyhatts]

need some help. whenever i try to access stable diffusion thru command prompt i always get this error. when i press any key the cmd prompt just disappear. im using comfyui

Not sure about the errors you are seeing.
But I would delete those existing standalone installations of SD-webui and ComfyUI.
And reinstall them using Stability Matrix.

 

[doogyhatts]

AI猫咪 trend.

 

[zheng]

 

[doogyhatts]

 

[doogyhatts]

Missed out this other project from ByteDance.

ResAdapter: Domain Consistent Resolution Adapter for Diffusion Models​

 

[doogyhatts]

Recent update to Automatic-CFG shows good hands.

 

[AZE]

New paper from NVidia!
Align Your Steps


Looks like they fixed SVD as well with this paper.
However, after testing it again, SVD still performs worse than DynamiCrafter.

https://stable-diffusion-art.com/align-your-steps/

Not much difference from Karras schedule according to the review.
But as I hab already suspected based on the noise schedule,

The heavier denoising at the start will always destroy the high frequency data(basically small details and textures) a lot more, and as shown by the article but not mentioned by the author, the pictures from AYS are sometimes more washed out, paintery and pastel rike and hab lower dynamics. Meaning will rikely negatively affect rendition of realistic pics/photos.

 

[doogyhatts]

Longer consistent videos can be generated soon using Story Diffusion.