I did quite a long series of tests a while ago where I took them out one by one. I found four was about the maximum as they weaken each other when too many are used. It also made a difference if the keywords had some relevance to the prompt subject. Try Baroque or Rococo they are equally as strong in some circumstances.
You are right, most do not know what they are doing. And the keywords come from the image captioning process, some model creators do not know much about photography.
Very interesting. I completely agree about avoiding the use of "magic" prompt words as much as possible but some magic words do seem to produce very predictable holistic improvements. The same goes for certain "magic" negative prompts. Mateo of IP Adapter fame has pointed out that negative prompts like "horror", "vampire" or "werewolf" push the model in a more "brighter" direction. I don't use them by default but will occasionally throw them in to give an image I'm working on a kick that direction. I'll be adding some of these to my positive prompting quiver as well. Thank you.
Interesting, for the me the ones without the keyword looked more natural, the ones with the keywords looked like like they had been retouched too much to make certain aspects standout.
thanks for the keyword list. the problem that lots of checkpoints seem to have with light and shadows are the fragments displayed on the model, like its standing under a umbrella with lots of holes in it and the sun is pointing those spots on the model. sadly, i have not found any negative prompts that could remove this besides changing the checkpoint
I haven't tested but in general the consensus is that (key)words at the beginning of a prompt get a higher emphasis than at the end. I never use long prompts so I never run out of tokens.
Octane render is a 3d render engine, it's one of the most physically correct in terms of how light behaves. Probably SD is trained an a lot of 3d artists renders done on Octane.
Yeah, like Vray, and Unreal engine, and POVray, one of the first renderers 40 years ago. Lol, I wonder if POVray is known in SD ... I'll give it a try. :)
Thank you for this vid 👍. I have question that hopefully someone can answer: I just upgraded to a new laptop with RTX 4090 16GB. Now my comfyUI monitor (Crystools node) shows VRAM usage at 99%, and GPU usage at 0%. 1) is this how it's supposed to be? 2) if not, what can I do to have the GPU run as well? Thanks to anyone that can help me with this.
Preference for a image or in this case style doesn’t always mean better…also the use of such keywords are situational and their use should be based on where final destination of the image as the keywords to skew the image towards a stylized and unnatural presentation. For example, if i plan to generate an image for a face to use as a texture or some other workflow, then the keywords distort and possibly “color” my final output. If however, you are going straight from the generated image to final output…then using these keywords can help in acting like a post-processing effect. All the images you presented, the ones on the left, the ones without keywords, looked natural, very close to traditional photography, while the keyword versions looked “AI Generated”. In cases of the stylized image (oil painting) then the keyword version was always better and I think should always be used when generating stylized or digital art subjects…otherwise, it just looks unnatural. Prompt stuff sucks…but because of models being trained on these keywords, only because people key are using them…we may never see this stuff go away
These “keywords” have been known for a very long time, and I also agree with you about disliking them. So, I also rarely use them because I prefer my work NOT to look like AI art. I forget which UA-camr was selling style packs last year, and all of those 15 keywords were in the “default” style selection, with a couple of the other ones like Greg Rutkowski. Note: AI is going to kick our butts because we always look for shortcuts, are predictable, lazy and are more like sheep than even sheep.
So very grateful for your research!!
As a vegetable bird for SD, I'm so excited to watch this video. The low-quality image confused me for a long time. It indeed helps me a lot.
Thanks for your feedback. Nice that it helps you. You might also like to see videos 26, 27, 28, they can help te get even nicer images.
Great video. Great information. Great voice. Thanks
I did quite a long series of tests a while ago where I took them out one by one. I found four was about the maximum as they weaken each other when too many are used. It also made a difference if the keywords had some relevance to the prompt subject. Try Baroque or Rococo they are equally as strong in some circumstances.
You are right, most do not know what they are doing. And the keywords come from the image captioning process, some model creators do not know much about photography.
Can u name the models whose creators know well about photography?
many many thanks, very helpful!
You're welcome!
Very interesting. I completely agree about avoiding the use of "magic" prompt words as much as possible but some magic words do seem to produce very predictable holistic improvements. The same goes for certain "magic" negative prompts. Mateo of IP Adapter fame has pointed out that negative prompts like "horror", "vampire" or "werewolf" push the model in a more "brighter" direction. I don't use them by default but will occasionally throw them in to give an image I'm working on a kick that direction. I'll be adding some of these to my positive prompting quiver as well. Thank you.
Interesting, for the me the ones without the keyword looked more natural, the ones with the keywords looked like like they had been retouched too much to make certain aspects standout.
Thank you for these keywords it help a lot !!
Glad it was helpful!
thanks for the keyword list.
the problem that lots of checkpoints seem to have with light and shadows are the fragments displayed on the model, like its standing under a umbrella with lots of holes in it and the sun is pointing those spots on the model.
sadly, i have not found any negative prompts that could remove this besides changing the checkpoint
Does it matter hte order of the keywords inm your testing ? Are you running out of tokens and then does it matter the order for logner prompts?
I haven't tested but in general the consensus is that (key)words at the beginning of a prompt get a higher emphasis than at the end. I never use long prompts so I never run out of tokens.
good job :) thanks keep going
Thanks for your feedback.
Octane render is a 3d render engine, it's one of the most physically correct in terms of how light behaves. Probably SD is trained an a lot of 3d artists renders done on Octane.
Yeah, like Vray, and Unreal engine, and POVray, one of the first renderers 40 years ago. Lol, I wonder if POVray is known in SD ... I'll give it a try. :)
thanks a few of them i was already using incorporating the rest i noticed a better generation, and for negatives any recommendations?
Well ... I seldom use negatives. Dreamshaper Turbo, which I now use 80% of the time, does not need much in that respect.
The images looks more ehmm Artistic and elaborated. I would said is actually an improved version.
Thank you for this vid 👍.
I have question that hopefully someone can answer:
I just upgraded to a new laptop with RTX 4090 16GB. Now my comfyUI monitor (Crystools node) shows VRAM usage at 99%, and GPU usage at 0%.
1) is this how it's supposed to be?
2) if not, what can I do to have the GPU run as well?
Thanks to anyone that can help me with this.
Does an image come out after say 10seconds when you use Dreamshaper Turbo at 6 steps? If so, All is OK.
4090 has 24GB VRAM
Preference for a image or in this case style doesn’t always mean better…also the use of such keywords are situational and their use should be based on where final destination of the image as the keywords to skew the image towards a stylized and unnatural presentation. For example, if i plan to generate an image for a face to use as a texture or some other workflow, then the keywords distort and possibly “color” my final output. If however, you are going straight from the generated image to final output…then using these keywords can help in acting like a post-processing effect. All the images you presented, the ones on the left, the ones without keywords, looked natural, very close to traditional photography, while the keyword versions looked “AI Generated”.
In cases of the stylized image (oil painting) then the keyword version was always better and I think should always be used when generating stylized or digital art subjects…otherwise, it just looks unnatural. Prompt stuff sucks…but because of models being trained on these keywords, only because people key are using them…we may never see this stuff go away
All very true.
These “keywords” have been known for a very long time, and I also agree with you about disliking them. So, I also rarely use them because I prefer my work NOT to look like AI art. I forget which UA-camr was selling style packs last year, and all of those 15 keywords were in the “default” style selection, with a couple of the other ones like Greg Rutkowski. Note: AI is going to kick our butts because we always look for shortcuts, are predictable, lazy and are more like sheep than even sheep.
90℅ of the people who use the keyword Greg Rutowski havent seen his art 😂
Now I’m tempted to look him up. I’ll try prompting with his name first