Stable diffusion thread

[doogyhatts]

Meow Motion remains the top AI creator for YT shorts that caters to both gender.

Spoiler

 

[doogyhatts]

More model creators have begun fine-tuning their datasets for PixArt-Sigma.

 

[zheng]

More model creators have begun fine-tuning their datasets for PixArt-Sigma.

i saw got pixart A and E got what difference?

 

[doogyhatts]

i saw got pixart A and E got what difference?

A is alpha, E is Sigma.
alpha is the initial release, sigma is the upgraded version.

@AikiBoy
Now that you understood how to use ComfyUI,
you can get output quality closer to Midjourney if you use PixArt-Sigma. Use this workflow.
Prompt adherence is not 100% though, so need to do inpainting using Fooocus.

 

[doogyhatts]

As for Lumina, have to wait until they finish researching the 16-channel VAE that SD3 uses.
In the meantime, Kijai has provided a ComfyUI wrapper for it.
Maybe worth trying out their Next-SFT model first and the Gemma-2b LLM that it uses.

 

[AikiBoy]

A is alpha, E is Sigma.
alpha is the initial release, sigma is the upgraded version.

@AikiBoy
Now that you understood how to use ComfyUI,
you can get output quality closer to Midjourney if you use PixArt-Sigma. Use this workflow.
Prompt adherence is not 100% though, so need to do inpainting using Fooocus.

i still learning all the menus/managers.

all the programming terms... see liao head faint.

What is VAE what is conditioning this and that. how to link nodes...

 

[AikiBoy]

Meow Motion remains the top AI creator for YT shorts that caters to both gender.

Spoiler

can monetise liao?

 

[tokong]

can monetise liao?

in youtube you need 1000 subscribers to start monetizing. i am recently creating my own dalle3 slideshow on youtube too like doggyhats, must slowly build up one. i now only have 313 subscribers.

 

[AikiBoy]

in youtube you need 1000 subscribers to start monetizing. i am recently creating my own dalle3 slideshow on youtube too like doggyhats, must slowly build up one. i now only have 313 subscribers.

sic your channel lol

i think ppl here must support each other channel lol.

 

[focus1974]

in youtube you need 1000 subscribers to start monetizing. i am recently creating my own dalle3 slideshow on youtube too like doggyhats, must slowly build up one. i now only have 313 subscribers.

photos generated by the checkpoints which are from SD1.5 to sdxl .. can use for commercial use meh?
monetizing is already commercial ?

 

[doogyhatts]

photos generated by the checkpoints which are from SD1.5 to sdxl .. can use for commercial use meh?
monetizing is already commercial ?

He used Bing-Dalle3.
Monetising YT channel is considered commercial.
Custom SD1.5 & SDXL models should be ok since they are not derivative works based on their respective base models.

 

[AikiBoy]

photos generated by the checkpoints which are from SD1.5 to sdxl .. can use for commercial use meh?
monetizing is already commercial ?

He used Bing-Dalle3.
Monetising YT channel is considered commercial.
Custom SD1.5 & SDXL models should be ok since they are not derivative works based on their respective base models.

how they know if the picture you used are generated or not.
so many AI pic video on youtube... how to trace the picture back to the app you used...
the picture got hidden codes, you throw into the ai app they know its their?

 

[doogyhatts]

how they know if the picture you used are generated or not.
so many AI pic video on youtube... how to trace the picture back to the app you used...
the picture got hidden codes, you throw into the ai app they know its their?

YT has a section that you have to declare your video content is AI-generated under certain circumstances.
This is mainly to prevent impersonations.

And you have to state what you use to create the content and how it is done in the description.
This is to allow the YT staff to understand your channel better when you apply for monetisation.

 

[tokong]

photos generated by the checkpoints which are from SD1.5 to sdxl .. can use for commercial use meh?
monetizing is already commercial ?

yup, microsoft says ok

https://learn.microsoft.com/en-us/a...he-images-generated-by-openai-dalle-commercia

 

[AikiBoy]

yup, microsoft says ok

https://learn.microsoft.com/en-us/a...he-images-generated-by-openai-dalle-commercia

which to dalle now!!

 

[focus1974]

He used Bing-Dalle3.
Monetising YT channel is considered commercial.
Custom SD1.5 & SDXL models should be ok since they are not derivative works based on their respective base models.

i think most of the checkpoints are derivatives of the base models.
they used the base model and then just input a few hundred or thousands photos of their own to train.

 

[AikiBoy]

i think most of the checkpoints are derivatives of the base models.
they used the base model and then just input a few hundred or thousands photos of their own to train.

chim, what hundred of photos to train?
if i dl a model, i keep using it, it become better? lol

 

[focus1974]

chim, what hundred of photos to train?
if i dl a model, i keep using it, it become better? lol

so what you are using is already trained with custom photos on the base model.
This is the typical process of training a checkpoint. You dont trained it, because you are already downloading the trained model from civitai or other sources.

Yes, checkpoints on Civitai are typically trained from the base Stable Diffusion (SD) model as the starting point. Here are the key points:

• The base SD model (e.g. SD 1.5, SD 2.0) serves as the foundation or "checkpoint" that is further fine-tuned or trained on specific datasets to create new specialized checkpoints.
• Training a checkpoint involves taking the base SD model weights and continuing the training process on a new dataset of images to steer the model towards a particular style, subject matter, or capability.
• This allows creators to specialize the general base SD model into more focused checkpoints tailored for certain artistic styles (e.g. anime, photorealism), subjects (e.g. portraits, landscapes), or other custom training objectives.
• Using the base SD checkpoint as the starting point allows the specialized training to build upon the broad knowledge and capabilities already learned by the original SD model.
• Some creators may use other pre-trained checkpoints besides the official SD releases as their base, but the common practice is to start from one of the official SD checkpoints and fine-tune from there.

So in summary, while techniques like LoRA allow modifying the SD model in other ways, the typical workflow for creating new custom checkpoints on Civitai involves fine-tuning the base SD model through further training on curated datasets.

 

[AikiBoy]

so what you are using is already trained with custom photos on the base model.
This is the typical process of training a checkpoint. You dont trained it, because you are already downloading the trained model from civitai or other sources.

So is like saying SD1.5 2.0 is the base.

like playing MMO rpg. you train the character on a certain job... you boost the level to the max, you add different stats points along the way. When strong enough for endgame content. You sell it or give it away. the next person who get your character cannot train anymore because it max stats.

 

[doogyhatts]

i think most of the checkpoints are derivatives of the base models.
they used the base model and then just input a few hundred or thousands photos of their own to train.

Hmm... I see.
Actually, no need to worry since you are not creating a new model to sell based on SD core models.
The license agreement states outputs from the core models are not derivative works, so it is only the custom models that the license is concerned with.