[Stable Diffusion] Prompt Sharing and Learning Thread

[Fuchsschweif]

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.

 

[rayminator]

I am getting this error from sd-webui-additional-networks is there something else i can use

*** Error running process_batch: C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\additional_networks.py
    Traceback (most recent call last):
      File "C:\Users\raymi\AI Images\webui\modules\scripts.py", line 864, in process_batch
        script.process_batch(p, *script_args, **kwargs)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\additional_networks.py", line 244, in process_batch
        network, info = lora_compvis.create_network_and_apply_compvis(
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 153, in create_network_and_apply_compvis
        network = LoRANetworkCompvis(text_encoder, unet, multiplier_tenc, multiplier_unet, modules_dim, modules_alpha)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 341, in __init__
        comp_vis_lora_name = LoRANetworkCompvis.convert_diffusers_name_to_compvis(self.v2, du_lora_name)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 302, in convert_diffusers_name_to_compvis
        assert cv_name is not None, f"conversion failed: {du_name}. the model may not be trained by `sd-scripts`."
    AssertionError: conversion failed: lora_unet_double_blocks_0_img_attn_proj. the model may not be trained by `sd-scripts`.

 

[Ashira13]

I am getting this error from sd-webui-additional-networks is there something else i can use

*** Error running process_batch: C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\additional_networks.py
    Traceback (most recent call last):
      File "C:\Users\raymi\AI Images\webui\modules\scripts.py", line 864, in process_batch
        script.process_batch(p, *script_args, **kwargs)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\additional_networks.py", line 244, in process_batch
        network, info = lora_compvis.create_network_and_apply_compvis(
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 153, in create_network_and_apply_compvis
        network = LoRANetworkCompvis(text_encoder, unet, multiplier_tenc, multiplier_unet, modules_dim, modules_alpha)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 341, in __init__
        comp_vis_lora_name = LoRANetworkCompvis.convert_diffusers_name_to_compvis(self.v2, du_lora_name)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 302, in convert_diffusers_name_to_compvis
        assert cv_name is not None, f"conversion failed: {du_name}. the model may not be trained by `sd-scripts`."
    AssertionError: conversion failed: lora_unet_double_blocks_0_img_attn_proj. the model may not be trained by `sd-scripts`.

looks like you need to update the extension to a new version. seen reports like that on github before.

 

[rayminator]

looks like you need to update the extension to a new version. seen reports like that on github before.

i am using Version 0.5.0 and still getting that error

 

[Ashira13]

i am using Version 0.5.0 and still getting that error

did you try updating

You must be registered to see the links

to version 0.8.7? looks like sd-webui-additional-networks hasn't been updated in over a year.

 

[Ashira13]

i am using Version 0.5.0 and still getting that error

you could try switching over to

You must be registered to see the links

,might test it out myself, when i upgrade my graphics card. 8gb vram is a bit low, even for generating images.

 

[Sharinel]

Thanks! So the first result turned out pretty nice:


View attachment 4050767

But the final result is messy:

View attachment 4050766

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.

What are you trying to do with it? Were you just trying to double it in size? I had a quick look at your workflow but as I don't use comfy flowcharter tab I don't know much about it. I did have a stab at it though with this model -

You must be registered to see the links

Here's what I got (hopefully the workflow is still there)

 

[Fuchsschweif]

What are you trying to do with it? Were you just trying to double it in size? I had a quick look at your workflow but as I don't use comfy flowcharter tab I don't know much about it. I did have a stab at it though with this model -

You must be registered to see the links

Here's what I got (hopefully the workflow is still there)

Yes I think it's doubling it in size. With Dreamshaper 8 I used the exact same settings, except that I increased the initial tile to 1024 (instead of 512) for this pony model since it was trained at 1024.

So I guess if I change the first one to 1024 I have to change something later on in the workflow too.. I just don't know what.

 

[Sharinel]

Yes I think it's doubling it in size. With Dreamshaper 8 I used the exact same settings, except that I increased the initial tile to 1024 (instead of 512) for this pony model since it was trained at 1024.

So I guess if I change the first one to 1024 I have to change something later on in the workflow too.. I just don't know what.

I think you are overthinking it. As you said, you changed the first one to 1024 then you have a node that upscales by 2x. What else do you need to do? That upscale node puts it to 2048 anyway? Or am I missing something?

Also have a look at the workflow I added with that picture, does that work on your end?

 

[rayminator]

I am getting this error from sd-webui-additional-networks is there something else i can use

*** Error running process_batch: C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\additional_networks.py
    Traceback (most recent call last):
      File "C:\Users\raymi\AI Images\webui\modules\scripts.py", line 864, in process_batch
        script.process_batch(p, *script_args, **kwargs)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\additional_networks.py", line 244, in process_batch
        network, info = lora_compvis.create_network_and_apply_compvis(
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 153, in create_network_and_apply_compvis
        network = LoRANetworkCompvis(text_encoder, unet, multiplier_tenc, multiplier_unet, modules_dim, modules_alpha)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 341, in __init__
        comp_vis_lora_name = LoRANetworkCompvis.convert_diffusers_name_to_compvis(self.v2, du_lora_name)
      File "C:\Users\raymi\AI Images\webui\extensions\sd-webui-additional-networks\scripts\lora_compvis.py", line 302, in convert_diffusers_name_to_compvis
        assert cv_name is not None, f"conversion failed: {du_name}. the model may not be trained by `sd-scripts`."
    AssertionError: conversion failed: lora_unet_double_blocks_0_img_attn_proj. the model may not be trained by `sd-scripts`.

does anyone else have a suggestion that i can use

i found this one Lora Queue Helper for SD WebUI A1111

someone else is having the same problem

You must be registered to see the links

You must be registered to see the links

 

[felldude]

You must be registered to see the links

THE FP8 MODEL IS WORKING BETTER THEN FP16 at least with the FP8 T5xxl, not tested with FP16 T5xxl.

• I posted a FP16 version, and have a BF16 yet to be released.
  100's of hours of training went into this model see pro's/cons.
  
  Pros
• Faster then GGUF and NF4 in my testing by 20%-40% (FP8 Version)
• Higher details then Q4_0 or NF4 Dev
• Higher details and more accurate then any LORA with the Q4_0 model in my testing.
• Ages 18-40yo trained on models with 2257 forms.
• Lora's are working with the FP8 quantized version (Comfy Workflows on my image)

Cons

• Sometimes a censored image will still slip through
• Ethnicity information was not trained and this seems to have caused an issue. This model will favor light skinned models to an extent I have not seen before.
• While Lora's work with FP8 they can add 30-50% more time to render
  

My Speed on a 3050 8GB
FP8

• 832x1216 @ 5.5 Seconds per IT
• 1024x1024 @ 6 Seconds per IT
• 1024x1536 @ 15 Seconds per IT
• 2048x2048 @ 28 Seconds per IT

FP16/BF16

• 832x1216 @ 8 seconds per IT
• 1024x1024 @ 10 seconds per IT
• 1024x1536 @ 11 seconds per IT (?) Why is it faster then FP8 maybe the cfg?

This model is a training using many individuals with known ages and 2257 forms, it has also been merged to try and ensure that no known individuals can be reproduced. However FLUX seems to like to learn faces even with less then 10% data rather then merge them into a new face.



BONUS
Based off of no other information then tons of testing, I think flux used random objects to censor nudity and explicit images. This makes it very hard to UN-censor as things like pillows, flowers and candy get super imposed over the image.



Bikinis combined with random objects based of unguided image to image testing.

 

[felldude]

More examples from

You must be registered to see the links

 

[Synalon]

Does it work in Forge or is it Comfy only?

 

[felldude]

Does it work in Forge or is it Comfy only?

If that is in reference to the model I posted, I have had reports of issues with forge where it shows it rendering in preview but black image at save.

It should work in forge with the correct vae, I updated the page with links

Also a few more pics using a exhibitionism lora and a black female character lora

 

[devilkkw]

felldude wow. I see you are working very well with flux.
Nice work.

 

[felldude]

felldude wow. I see you are working very well with flux.
Nice work.

Thank you

 

[felldude]

You must be registered to see the links

A heads up I have over 25,000 worth of buzz (25$) up for grabs over at

You must be registered to see the links

 

[TheREALMorlock]

I'm a total noob with genAI (I have a 4gb Radeon card so I haven't been motivated to dive into local solutions); is there a way to take an image created in Daz Studio (or Blender) and upres it with AI to make it look as good as the stuff people are posting here, without changing the overall composition of the original?

Edit:
I just saw this:
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-14128769

 

[Sharinel]

I'm a total noob with genAI (I have a 4gb Radeon card so I haven't been motivated to dive into local solutions); is there a way to take an image created in Daz Studio (or Blender) and upres it with AI to make it look as good as the stuff people are posting here, without changing the overall composition of the original?

Edit:
I just saw this:
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-14128769

Img2img does this to an extent. You can feed it an image from whatever source you want, ask it to apply a prompt to it and depending on the strength of your denoising you can tweak the output.

Example - Here's a pic I took from one of my fav VN's - The Intoxicating Flavor



I added this to img2img, and tried to mimic it with the prompt :-
This is a photorealistic photo depicting a young woman standing in a modern bathroom. She has fair skin and long, wavy brown hair that cascades down her shoulders. Her facial features are detailed, with high cheekbones and full lips. She wears a light blue, ruffled tank top that is slightly lifted, revealing a glimpse of her midriff and a small, silver belly button ring. Her shorts are blue, distressed denim, and fit snugly around her hips, showcasing her slim, toned legs. The bathroom is minimalist and modern, featuring a light gray brick wall in the background, which adds texture and a rustic feel. There is a large mirror to the left, reflecting part of the scene, including the woman's reflection and the gray brick wall. The floor is a light, polished stone, adding to the sleek and contemporary aesthetic. A white, smooth countertop extends into the foreground, where the woman's hand rests. The lighting is soft and natural, suggesting daylight filtering through the room. The overall mood is casual and relaxed, capturing a moment of everyday life in a modern home.

Here was the outcome at 0.4 strength denoising


As you can see not much has changed apart from the face. Now at 0.7


Much more realistic face but the AI is beginning to add it's own spin - look at the reflection changing in the mirror.

If you want to go nuts, put the denoise up to 0.9. It breaks away a lot more from following the pic directly and adds more of the underlying model to the image instead ( In this case Flux dev)


And that is just img2img. You have inpainting and a lot of different controlnets you can use, but as I don't have the first clue about using them I'll let someone else explain

 

[rogue_69]

Img2img does this to an extent. You can feed it an image from whatever source you want, ask it to apply a prompt to it and depending on the strength of your denoising you can tweak the output.

Example - Here's a pic I took from one of my fav VN's - The Intoxicating Flavor

I've been using the flux inpaint to make the faces look more realistic. I've created a Flux Lora for the face I want, used that to create a bald image of that face (using img2img from a bald Daz render with flat lighting), and used that in FaceGen to get the Daz face really close to the Lora. Then, I just use Flux on the face, making it look a little more lifelike. The good thing too, is that if you do a short animation, and only inpaint the face, you don't get much flickering (hair is a problem with img2img animations, as well as wandering torso freckles, moles, and nipples moving on you.) For me, I'm just trying to get the face to look less plastic so, I just keep the denoise around 0.25, which keeps it consistent. First image Daz, second is 0.25 denoise, last is 0.85 denoise with a smile added to prompt.