Stable diffusion thread

[zheng]

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

still got no public release of SD3... think sai going bankrupt liao

 

[doogyhatts]

I haven't tried PixArt yet. It uses T5 encodings instead of Clip.

Hmm... Nerdy Rodent says he could use it within 6gb vram.
Unlikely he loaded the 2K model as the model was released a few days later after his video was uploaded.

Abominable Spaghetti Workflow v2 was released on 10th May, so its not too late to upgrade.

 

[zheng]

Hmm... Nerdy Rodent says he could use it within 6gb vram.
Unlikely he loaded the 2K model as the model was released a few days later after his video was uploaded.

haiz comfyui

 

[doogyhatts]

It turns out that I can generate PixArt-Sigma images after all!
But the T5 encoder is very slow, especially if the prompt gets changed.
So it should only be used when the prompts are finalised.

The wide angle shot now works because the Abominable Spaghetti workflow is specially made unlike the online demo, and uses the Photon (SD1.5) model as the refiner.

 

[doogyhatts]

Trying to see if it can generate the Kappa yokai.
Couldn't generate the water dish on the head and the turtle carapace on its back is missing.

 

[doogyhatts]

ToonCrafter.
Similar to DynamiCrafter, but for cartoon style images.

 

[doogyhatts]

Just upgraded to use a more compact version of the t5 model.
Saves 9.5gb disk space and t5 encoding is now much faster at 5 minutes.

I had to shutdown other browsers (except the one with ComfyUI) and all other non-essential applications, since it uses most of the system memory, leaving 0.5 to 1.0 gb available.
Since I am using the cpu encoding, it might be a lot faster for those with 24gb vram and use gpu to do the encoding instead.

 

[Scarlet Tea]

What base model is good for generating horror themed images like ghost or demons?

 

[doogyhatts]

What base model is good for generating horror themed images like ghost or demons?

Depends on what type of ghost or demons.
The base model is shown for the images you see on Civit.ai.
So I normally check out what kinds of LoRAs are available for ghosts and demons, and then check out their base model, settings and prompts.

 

[focus1974]

what checkpoints are good for chinese models?

 

[zheng]

what checkpoints are good for chinese models?

china chinese - guofeng.... ronghua.... only know this two popular one...

otherwise leosam.... majic....

there are many more can search for it

 

[focus1974]

guofeng.... ronghua.... only know this two popular one... there are many more can search for it

ok. civitai.
i tried again

 

[doogyhatts]

He is actually adding RPG-Diffusion.

Output from their HF demo which uses Realistic Vision 4 XL.
Hands still problematic but cheongsam floral print is close to Dalle3 standard.
Face can be fixed using ADetailer or do face swap using ReActor.
The auto-generated sentences are very useful in re-writing prompts and making them a lot more detailed.

Not sure how to configure to use another model as the model filenames are different on HF.
I think both Forge and Fooocus will get this integration since they are all created by him.

 

[doogyhatts]

Instance Diffusion with ComfyUI support.
This allows for more precise control to image generation by breaking up the image into smaller chunks.
The example videos showed movement for a crowd of entities.

Next, tried IC-light HF just now.
The HF site is easier to use and I got better results, because the ComfyUI examples are not the same as the online demo.
Very useful in getting different backgrounds and directional relighting.

Below is the last keyframe of an existing animation generated using DynamiCrafter. So the hands degraded.
Speaking of which, the DC wrapper by Kijai now has ToonCrafter support as well.

 

[doogyhatts]

@zheng
The weights for SD3 Medium will be released on 12th June.

 

[doogyhatts]

 

[zheng]

Instance Diffusion with ComfyUI support.
This allows for more precise control to image generation by breaking up the image into smaller chunks.
The example videos showed movement for a crowd of entities.

Next, tried IC-light HF just now.
The HF site is easier to use and I got better results, because the ComfyUI examples are not the same as the online demo.
Very useful in getting different backgrounds and directional relighting.

Below is the last keyframe of an existing animation generated using DynamiCrafter. So the hands degraded.
Speaking of which, the DC wrapper by Kijai now has ToonCrafter support as well.

the relight look kinda unnatural

 

[AikiBoy]

It turns out that I can generate PixArt-Sigma images after all!
But the T5 encoder is very slow, especially if the prompt gets changed.
So it should only be used when the prompts are finalised.

The wide angle shot now works because the Abominable Spaghetti workflow is specially made unlike the online demo, and uses the Photon (SD1.5) model as the refiner.

new software again?

wao, i so outdated liao. what is good to use now?

 

[doogyhatts]

new software again?

wao, i so outdated liao. what is good to use now?

It is ComfyUI, quite complex for beginners.
SD3 will be released on 12th June.

Omost is the easiest to use for beginners right now. But it is limited to Realistic Vision 4 XL model right now. Can also wait for the China expert to upgrade it to SD3.

For animation, Haiper is still free right now.

 

[AikiBoy]

It is ComfyUI, quite complex for beginners.
SD3 will be released on 12th June.

Omost is the easiest to use for beginners right now. But it is limited to Realistic Vision 4 XL model right now. Can also wait for the China expert to upgrade it to SD3.

For animation, Haiper is still free right now.

installing the app look complicated...