[Stable Diffusion] Prompt Sharing and Learning Thread

[felldude]

You must be registered to see the links

Tested at 1280x1536 locally on 8GB video card, using the

You must be registered to see the links

NF4 is still faster for me but the lack of LORA support makes it useless.

This image took me 263 seconds on a RTX 3050 (20 Iterations)

 

[felldude]

Example Testing multiple LORA at near 2k using image to image as a BLIP integration prompt

 

[felldude]

You must be registered to see the links

This is a dataset prepared for FLUX assuming Civitai adds TE training
But I tested it via PONY, along with my other BLIP-2 model it has a fair grasp of natural language.

 

[Deleted member 1666680]

Stable Diffusion seems to be stuck in a progress-crisis. After the first initial huge leaps forward there's not much happening right now.. we still need too many external things to get basic stuff done.

 

[Ashira13]

An example of working with SDXL Lightning checkpoints with refiner and various Loras plus HiresFix. : Model :

You must be registered to see the links

, Refiner :

You must be registered to see the links

 

[felldude]

Stable Diffusion seems to be stuck in a progress-crisis. After the first initial huge leaps forward there's not much happening right now.. we still need too many external things to get basic stuff done.

It comes down to basic math, you take the number of parameters by the data storage type.
Without quantization and data parallelism we wouldn't be using the size models we are.

No one has bothered to do it but a NF4 or Q4_0 model of SDXL should be around 2-3GB

 

[Deleted member 1666680]

It comes down to basic math, you take the number of parameters by the data storage type.
Without quantization and data parallelism we wouldn't be using the size models we are.

No one has bothered to do it but a NF4 or Q4_0 model of SDXL should be around 2-3GB

(I have no idea what you're talking about)

 

[felldude]

View attachment 4044151

(I have no idea what you're talking about)

We are reaching the limits of reducing the blocks size down without massive data loss.

Chat GPT is thought to have 1 trillion parameters without some form of reducing the size that would take 14 Terabytes of VRAM to train

 

[Deleted member 1666680]

We are reaching the limits of reducing the blocks size down without massive data loss.

Chat GPT is thought to have 1 trillion parameters without some form of reducing the size that would take 14 Terabytes of VRAM to train

That's really cool! Meanwhile my Stable Diffusion can't do hands or feet properly, no penetration, leave alone two normally functioning people in one image, and when I specify I want a "ghost girl" I always have bedsheet-ghosts in the background.

"Hairy pussy" often ends up with a complete half-human-half-ape creature.

I can't wait until Stable Diffusion finds out about its technological progress

 

[felldude]

That's really cool! Meanwhile my Stable Diffusion can't do hands or feet properly, no penetration, leave alone two normally functioning people in one image, and when I specify I want a "ghost girl" I always have bedsheet-ghosts in the background.

"Hairy pussy" often ends up with a complete half-human-half-ape creature.

I can't wait until Stable Diffusion finds out about its technological progress

One of the many finetunes of PONY would likely be a better fit for what your asking.

 

[Deleted member 1666680]

One of the many finetunes of PONY would likely be a better fit for what your asking.

Turned out even worse. Any idea why?

 

[felldude]

Turned out even worse. Any idea why?

you only have it set at 0.6 your up-scaling a 512x512, pony works best at 1024x1024
Your trying to use XL vae with PONY

 

[Deleted member 1666680]

you only have it set at 0.6 your up-scaling a 512x512, pony works best at 1024x1024
Your trying to use XL vae with PONY

I tried 1024x1024 too, didn't get better.. and the VAE is the one that was directly linked on the page of this SDXL Pony 6 checkpoint. Is it wrong? Here's the 1024x1024 result.

 

[felldude]

I tried 1024x1024 too, didn't get better.. and the VAE is the one that was directly linked on the page of this SDXL Pony 6 checkpoint. Is it wrong? Here's the 1024x1024 result.

You might try one of the other PONY models that suites your taste.

This is my merge result same seed. With different prompts



Sexy Ghost is not an easy prompt though

 

[Deleted member 1666680]

You might try one of the other PONY models that suites your taste.

I usually work with Dreamshaper 8 which puts out good content. But I can' get Pony to run, it always looks as bad as my previous images. Obviously, something must be wrong with my settings but what?

 

[felldude]

I usually work with Dreamshaper 8 which puts out good content. But I can' get Pony to run, it always looks as bad as my previous images. Obviously, something must be wrong with my settings but what?

Load one of my images with any PONY model and see if the results get better

 

[Sharinel]

I usually work with Dreamshaper 8 which puts out good content. But I can' get Pony to run, it always looks as bad as my previous images. Obviously, something must be wrong with my settings but what?

Are you putting the score_9, score_8_up shit in the prompt? Otherwise it looks like a cartoon animal having a seizure

 

[Deleted member 1666680]

Are you putting the score_9, score_8_up shit in the prompt? Otherwise it looks like a cartoon animal having a seizure

No didn't do it, is it mandatory? Why all these different numbers?

 

[Sharinel]

No didn't do it, is it mandatory? Why all these different numbers?

Yeah it's mandatory. The absolute madlad who made the model intended it to be a complete rework of SDXL but was a bit too... intense. So it takes different inputs to really get it working good. Here's what I have as a style that I use with every pony pic

Additional prompts (paste at start of prompt) - score_9, score_8_up, score_7_up,
Addition negative prompts - score_6_up, score_5_up, score_4_up

He also used a prompting tag group called Danbooru, have a look here for the words to put in your prompts, especially for NSFW concepts.

You must be registered to see the links

Example - Here's a pic I just generated with the prompts above on a pony model


And here it is without any of the pony word-salad in the prompts/neg prompts. Same seed etc. Both of these are png files if you want to pop them in the png infor tab and have a look.

 

[Deleted member 1666680]

Yeah it's mandatory. The absolute madlad who made the model intended it to be a complete rework of SDXL but was a bit too... intense. So it takes different inputs to really get it working good. Here's what I have as a style that I use with every pony pic

Additional prompts (paste at start of prompt) - score_9, score_8_up, score_7_up,
Addition negative prompts - score_6_up, score_5_up, score_4_up

He also used a prompting tag group called Danbooru, have a look here for the words to put in your prompts, especially for NSFW concepts.

You must be registered to see the links

Example - Here's a pic I just generated with the prompts above on a pony model
View attachment 4049832

And here it is without any of the pony word-salad in the prompts/neg prompts. Same seed etc. Both of these are png files if you want to pop them in the png infor tab and have a look.

View attachment 4049834

Thanks! So the first result turned out pretty nice:




But the final result is messy:



Does this have to do with my tile_size in the final vae_decode?

Usually I use 512 as base so it's set to 512 too. But when I input 1024 now I get an error message.