I played with it all weekend and agree with you - Very good one. I even done a bunch of finetuning on it (over 3K image) and the results are pretty amazing!
I trained for a client of ours (Norma Kamali) on her swimwear archive, over 40yrs of archive! The results are so versatile and precise - Im in awe! @@sebastiankamph
From my experience. Simple prompts works better because the AI can "think more freely". Just like our brain, when we're told to many things to do, it's clouded and the output will be worse.
Civitai - A Civit is a small cat that lives in the rain forest and eats coffee berries that ferment in it's gut and its poo is the world's most expensive coffee
Great model, too bad it crashes with Deforum after generating around 200 images, see the following error message: User friendly error message: Error: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.. Please, check your schedules/ init values.
I am super perplexed by SDXL; its benefits seem to be tradeoffs, not upgrades. I basically learned 1.5 from your wisdom - huge thanks and praise, and I'd like to think I've gotten quite good at it, even on a 10y old potato. I notice the results of in-painting are weaker; you can't just force more pixel density into an area like 1.5. There seems to be this very common theme where blur and distortion are just accepted around background elements, where there is massive loss of detail for, dare I say, a more artistic look instead of something polished and sharp? I guess we are prioritizing foreground elements to go large and saying forget the rest... Between the refiner, vae, high res fix, and low support for ultimate upscale in a111, is it forcing you to jump to comfyui? I feel like this is a backwards progression. I have renders on 1.5 that are not even just as good; they are exponentially better. The one thing I do love about SDXL is the natural and very literal interpretations of prompting. Body position and senes do seem more dynamic and exciting, which I love, but dare I say I am tempted to start in SDXL and use it as an openpose to go back to 1.5 for my quality. You will likely get what you put in even without the major support of loras, so it feels like it has more scope. I just got a brand new PC with a GForce RTX 4070TI, and I thought I'd take the plunge and try the more intensive interface as SDXL was a 2 hour render and is now 14 seconds. Am I about to throw in the towel as it feels like way more trouble than its worth? Underwhelmed and kinda disapointed. Am I crazy, or does anyone else feel like this?! Perhaps, SDXL is a is a good idea on paper, but a different flavor of ice cream may not be the next big thing -idk maybe I am missing some important piece. I would be very curious what people's opinions are. With some great prompts and a couple good loras, you can drop a 3000x3000 image on a 1.5 set of dimensions to.3-.4 with a CFG of 7.5-10 and churn out how high the res masterpieces starting at 768x768, which may only need a little inpainting and minor upscaling. Running 768sx1344s regardless of the inputs I feel like I want to throw up in my mouth, getting results like your first day of stable diffusion without embeddings, loras, and using base models. It concerns me that the only refiner I seem to find for SDXL is the base model, which is universally garbage; running images through it seems to make things even worse, even at 80% or 70% shift. I am just wondering if anyone else feels the same? Just like 1.5 when you start to big it overwhelms the AI and by that standard I think SDXL and even the upcoming cascades are just flawed, start small get something amazing scale appropriately. When I look at top renders on CIVITAI a lot of the SDXLs don't really impress me, the hyper-realism is better ish, but all other styles just as good or worse, and often in far less quantity. Can 1.5 outperform SDXL as more veteran'd and polished system? I am curious to try for another week or two, and then I might just hold off or a year until it gets dialed in more. Thoughts and feedback welcome, really not sure what to think.
I tried this model and it is amazing. the 6 GB RTX 3060 on my gaming laptop takes a long time to render however. If I can get stable diffusion to see my desktop's RX 6800, might be different.
Love the videos, Sebastian, is there any chance of a longer form video where we see an video being initially made in Fooocus, and then touched up in automatic1111? Even so, this is definitely a model I'm going to try out ASAP.
I sometimes have that issue but it usually sticks if I wait about 10 seconds before trying again. I assume you've restarted the UI (including the console window) and so on? That also works for me. I know I'm stating the obvious but you never know. 🙂
Looking at the console/CMD is does say failed to load checkpoint, restoring previous. How do i see what version of SD this checkpoint uses?@@Elwaves2925
Be nice to see how this works for creating head/face with no hair for use with tools like face-transfer in Daz 3D and head-shot 2 in Chrataceter Creator
Guys do anyone have access to the Ruined Fooocus download file? After downloading and extracting, it is showing multiple critical errors, so it cant be even extracted properly from the zip folder
Is ruined focus faster at generating than AUTOMATIC 1111? I am a beginner & am wondering what my expectations should be on generating times. I have just started with Juggernaut SDXL.
RuinedFooocus will be a little faster for SDXL due to the fact that it's built upon Comfy and handles XL better. It's not a massive change, but a little bit.
Hi when I try to load this in A1111 it takes quite long then it doesn't load just simply goes back to the previous model , is it because my computer is too slow, how can I fix that
Try using a more lightweight ui like RuinedFooocus if your gpu is not up to the task. If you want to use SDXL with a1111, make sure you have the latest version.
@@sebastiankamph Okay, it definitely doesn't work. Trying to get a mermaid for a painting reference. (I always use simple prompts. No artist names, no loras, no styles.) It can't do hands, or noses, or eyes; the compositions are terrible; there are scales in the water and in the sky; it always ignores at least one part of my prompt. The skin texture is very realistic, but that's not helpful here at all. :-(
@@sebastiankamph CrystalClear is the only one that gives me consistent real-world accuracy across all prompt subjects - everything from figurines to futuristic cities to pools in desert palaces to trading cards and figurines, on and on. And I've tried close to fifty models. Most models work for certain subjects like: cars, faces, landscapes, and food. A few models work for fantasy like elves, dragons, and intricate armor. But all models except CrystalClear fail to follow prompts accurately across the subjects I need to paint, and I have no idea why!
same here, very difficult to get good stuff out of it, only thing I could get real good was joker's face makeup, almost nothing else, also tried the exact prompts with exact seed from the examples in civitai, surprisingly it still gives me awful stuff @@sebastiankamph
Indeed an awesome model. But a Q: under the "Performance" tab, how can I get that file like yours in this video? I mean all what I have to choose is "Speed", "Quality" and "Custom" which opens other options. But what is this "Seb_cfg4.." and how I get it?
I tried plenty of models, this one gives a pretty cinematic look but nothing can beat the speed of realisticvision, the speed and accuracy. Bulk image creation is too easy
@@sebastiankamph But what I'm asking is, isn't ThinkDiffusion XL using 1.0, which would make it not as good as a newer model? Why are still training models on the older version?
SORRY FOR being a unknowledgable european but can you tell me is there a local install for windiw 11 amd gpu to use this model on my pc ?! ive tried to get into AI art few montsh ago but there wasnt any way to get it to work because i have amd gpu .....
I can't speak for using an AMD GPU, it may cause issues but the official Automatic1111 install file has been available as standalone install file. No need to download any dependencies and so on, it does it all in one go. Maybe you last tried before that became a feature? Might be worth trying again. YT doesn't like links but it should be easy to find, or there's a link in the description for the one in the video, which is a good starter UI. 🙂
Not quite. Saying that it's a database of the images used for training implies that the images are stored in the model, but they aren't stored, the model merely learned from the images.
I modified the "webui-user.bat" and added the following arguments after set COMMANDLINE_ARGS: --xformers --medvram --no-half-vae. It helped with my img generation speed a lot (on rtx 2080ti, 11gb vram)
Is it just me, or does the model have a propensity to stretch the neck for tall images? (With tall, I mean something like 1216x832, which is only 1.46:1, nothing extreme)
For me its the opposite, been a photographer for 15 years and majority of models ive used give amazingly realistic results for portraits with the acception of occasional common body disfigurations with Ai. I find the challenge is finding a model that can do creative artworks specifically to your needs. A photograph i can be like it is what it is. But say like concept art, logo designs, character designs, etc. You need precise details to your vision. Otherwise its pointless to your project.
Had to delete it, with it as my checkpoint all my embeddings, and LORA's would disappear, and I did not have acess to them, when i went back to another checkpoint they came back, weird huh, anyway my favorite checkpoint is Lyriel, secound is dreamshaper
Do we really believe they hand captioned 10,000 images? Wouldn't that be like a full time job for a small team over several weeks. Combined with the fact they were also touting this as some kind of next level model on Reddit a few weeks back and then sponsoring UA-camrs makes me think they are planning on trying to convince people they have something special and then move future models to some kind of paid service. So far I'm yet to see anything from it that I couldnt get from some of the other XL models.
I've spoken to the people that did the work and they indeed say it was hand captioned. I see no reason to make that up, but I can also not verify that it's true. It's a cool feature that improves the model, but not revolutionary. As far as I know, the models are not planned to go paid and from a business perspective more of a marketing tool.
@@sebastiankamph Ok that's good to hear. I'm still skeptical though as hand captioning 10,000 images is a huge and extremely tedious task and not something you would expect done for nothing in return. Even at one caption a minute which is generous that's over 160 hours of mind numbing work.
Those millions of steps doesn't really show. Not much different than juggernaut. Actually worst in some cases for realistic results. The prompts shown here as examples are detailed in every model. Aliens in particular.
AIYou, I thought you've meant 3d models, which were made by humans, but no it's same hype of lazy stupid IA thing. IA will evolve human will degrade, not all, but many. Tools should help, not to replace human being at least in art.
Get early access to videos and help me, support me on Patreon www.patreon.com/sebastiankamph
I played with it all weekend and agree with you - Very good one. I even done a bunch of finetuning on it (over 3K image) and the results are pretty amazing!
Wow, what did you train on?
probably questionable shit ;) .. like it @@sebastiankamph
I trained for a client of ours (Norma Kamali) on her swimwear archive, over 40yrs of archive! The results are so versatile and precise - Im in awe! @@sebastiankamph
cool! love how there's so much free stuff in the opensource stable diffusion community
Awwww yeah, i just downloaded this model the other day and made it my defaults, and thankyou for showcasing RuinedFooocus
RuinedFooocus is fantastic, I love it! 🥰
Love trying new models, thanks for the recommendation
I got you! 😊🌟
The other 2 models that are up there with this one are _Mohawk_ and RelVisXL.
Oh, I haven't tried Mohawk, thanks for the tip!
I tried it, and the results are impressive! Thanks!
Great to hear! 🥰
It's surprising how such simple prompts you showed created such high quality images. Stable Diffusion has come a long way since just last year
yeah simple prompts work especially well with sd xl, sadly more specific prompts seem to be a little tricky to get remotely right.
Yeah, it's much closer to the simple prompting of Midjourney. Much easier to get good looking images quickly.
From my experience. Simple prompts works better because the AI can "think more freely". Just like our brain, when we're told to many things to do, it's clouded and the output will be worse.
Niiiiiiice my new go to! Thanks for sharing
In this video you said the prompt styles are free check out the link below. When I click on it wants me to sign up for a monthly plan????
Realism is nice, but I want better composition controls and better tools for changing specific details.
Bing Image generator(Dall.e 3,) is so much better at that. Hope it comes to SD soon.
Curious, what UI do you use?
ControlNet and InpaintAnything extensions on A1111 are my favorites right now for comp control...
For Stable Diffusion in general, that's already available in ControlNet and Inpainting. But everything can be improved I guess!
We missed you man :)
Oh, I'm around! 🥰
Great Bro, the model eas lit!!!
I've been using xeno for a certain style and juggernaut for more realism a lot, keen to try this out.
Let me know what you think! Would love to hear some more thoughts
what is xeno model ? what is its full name please ?
Really fantastic video, but I had to pause it over the umbrella comment 😂😂😂
10/10 quality, right? 🤣
Civitai - A Civit is a small cat that lives in the rain forest and eats coffee berries that ferment in it's gut and its poo is the world's most expensive coffee
Great model, too bad it crashes with Deforum after generating around 200 images, see the following error message:
User friendly error message:
Error: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.. Please, check your schedules/ init values.
For some reason this model runs super slow. Getting 6.4s/it when with most other SDXL models I get 10-15it/s
do u know why mine doesn't generate correctly? I get this weird highlighter red and yellow output... thanks
I am super perplexed by SDXL; its benefits seem to be tradeoffs, not upgrades. I basically learned 1.5 from your wisdom - huge thanks and praise, and I'd like to think I've gotten quite good at it, even on a 10y old potato. I notice the results of in-painting are weaker; you can't just force more pixel density into an area like 1.5. There seems to be this very common theme where blur and distortion are just accepted around background elements, where there is massive loss of detail for, dare I say, a more artistic look instead of something polished and sharp? I guess we are prioritizing foreground elements to go large and saying forget the rest... Between the refiner, vae, high res fix, and low support for ultimate upscale in a111, is it forcing you to jump to comfyui? I feel like this is a backwards progression. I have renders on 1.5 that are not even just as good; they are exponentially better. The one thing I do love about SDXL is the natural and very literal interpretations of prompting. Body position and senes do seem more dynamic and exciting, which I love, but dare I say I am tempted to start in SDXL and use it as an openpose to go back to 1.5 for my quality. You will likely get what you put in even without the major support of loras, so it feels like it has more scope. I just got a brand new PC with a GForce RTX 4070TI, and I thought I'd take the plunge and try the more intensive interface as SDXL was a 2 hour render and is now 14 seconds. Am I about to throw in the towel as it feels like way more trouble than its worth? Underwhelmed and kinda disapointed. Am I crazy, or does anyone else feel like this?! Perhaps, SDXL is a is a good idea on paper, but a different flavor of ice cream may not be the next big thing -idk maybe I am missing some important piece. I would be very curious what people's opinions are. With some great prompts and a couple good loras, you can drop a 3000x3000 image on a 1.5 set of dimensions to.3-.4 with a CFG of 7.5-10 and churn out how high the res masterpieces starting at 768x768, which may only need a little inpainting and minor upscaling. Running 768sx1344s regardless of the inputs I feel like I want to throw up in my mouth, getting results like your first day of stable diffusion without embeddings, loras, and using base models. It concerns me that the only refiner I seem to find for SDXL is the base model, which is universally garbage; running images through it seems to make things even worse, even at 80% or 70% shift. I am just wondering if anyone else feels the same? Just like 1.5 when you start to big it overwhelms the AI and by that standard I think SDXL and even the upcoming cascades are just flawed, start small get something amazing scale appropriately. When I look at top renders on CIVITAI a lot of the SDXLs don't really impress me, the hyper-realism is better ish, but all other styles just as good or worse, and often in far less quantity. Can 1.5 outperform SDXL as more veteran'd and polished system? I am curious to try for another week or two, and then I might just hold off or a year until it gets dialed in more. Thoughts and feedback welcome, really not sure what to think.
I tried this model and it is amazing. the 6 GB RTX 3060 on my gaming laptop takes a long time to render however. If I can get stable diffusion to see my desktop's RX 6800, might be different.
Pls make a work flow A-Z for this setting and all for A1111 pls
Im having an error, when i try to use this model in foooocus: AttributeError: 'NoneType' object has no attribute 'model'
Love the videos, Sebastian, is there any chance of a longer form video where we see an video being initially made in Fooocus, and then touched up in automatic1111? Even so, this is definitely a model I'm going to try out ASAP.
Sounds like a good idea!
Doesn't let me use it for some reason. I downloaded it but it just switches back to the model I was using previously.
I sometimes have that issue but it usually sticks if I wait about 10 seconds before trying again. I assume you've restarted the UI (including the console window) and so on? That also works for me. I know I'm stating the obvious but you never know. 🙂
Well it only does it for this model. Is that the case with you too? Maybe I don't have the right version of stable diffusion?@@Elwaves2925
Looking at the console/CMD is does say failed to load checkpoint, restoring previous. How do i see what version of SD this checkpoint uses?@@Elwaves2925
Be nice to see how this works for creating head/face with no hair for use with tools like face-transfer in Daz 3D and head-shot 2 in Chrataceter Creator
Guys do anyone have access to the Ruined Fooocus download file? After downloading and extracting, it is showing multiple critical errors, so it cant be even extracted properly from the zip folder
Is ruined focus faster at generating than AUTOMATIC 1111?
I am a beginner & am wondering what my expectations should be on generating times.
I have just started with Juggernaut SDXL.
RuinedFooocus will be a little faster for SDXL due to the fact that it's built upon Comfy and handles XL better. It's not a massive change, but a little bit.
Nice and thanks. Always willing to try out new models.
I say Civit-A-I because it's A.I. that we use.
Sometimes we are in a hurry and call it the other way! XD
Hi when I try to load this in A1111 it takes quite long then it doesn't load just simply goes back to the previous model , is it because my computer is too slow, how can I fix that
Try using a more lightweight ui like RuinedFooocus if your gpu is not up to the task. If you want to use SDXL with a1111, make sure you have the latest version.
Wow, this model creates incredibly realistic skin! I love how you've crafted such amazing images with it, I´m gonna try this out as well! Thx Seb!
Have fun! Let me know how it goes.
I was trying this over the weekend, and I couldn't get anything good out of it. I'll try it again.
Try it with simpler prompts, or if you're using a1111 my SDXL styles. Worked very well for me. The keyword cinematic helped a ton.
@@sebastiankamph Okay, it definitely doesn't work. Trying to get a mermaid for a painting reference. (I always use simple prompts. No artist names, no loras, no styles.) It can't do hands, or noses, or eyes; the compositions are terrible; there are scales in the water and in the sky; it always ignores at least one part of my prompt. The skin texture is very realistic, but that's not helpful here at all. :-(
I see, what XL models usually work for your usecase?@@jonmichaelgalindo
@@sebastiankamph CrystalClear is the only one that gives me consistent real-world accuracy across all prompt subjects - everything from figurines to futuristic cities to pools in desert palaces to trading cards and figurines, on and on. And I've tried close to fifty models. Most models work for certain subjects like: cars, faces, landscapes, and food. A few models work for fantasy like elves, dragons, and intricate armor. But all models except CrystalClear fail to follow prompts accurately across the subjects I need to paint, and I have no idea why!
same here, very difficult to get good stuff out of it, only thing I could get real good was joker's face makeup, almost nothing else, also tried the exact prompts with exact seed from the examples in civitai, surprisingly it still gives me awful stuff @@sebastiankamph
Hi thank you so much for this. I have a question how can we repose a character in an existing image?
Indeed an awesome model.
But a Q: under the "Performance" tab, how can I get that file like yours in this video? I mean all what I have to choose is "Speed", "Quality" and "Custom" which opens other options. But what is this "Seb_cfg4.." and how I get it?
If you go into custom and set your own settings, you can then save those custom settings.
I tried plenty of models, this one gives a pretty cinematic look but nothing can beat the speed of realisticvision, the speed and accuracy. Bulk image creation is too easy
Which model would you recommend for fantasy/dnd images,
This or Juggernaut probably
You shoud really fold your umbrella-jokes.
😂
ha ha😄
they are free for your patreons...
Basic question: I noticed that this model is using SDXL 1.0. Isn't 1.5 better?
Version numbers go in this order. 1.4 1.5 2.0 2.1 XL
@@sebastiankamph But what I'm asking is, isn't ThinkDiffusion XL using 1.0, which would make it not as good as a newer model? Why are still training models on the older version?
There is only one official SDXL base model. 1.5 is not SDXL and SD 1.5 is earlier than SDXL 1.0@@binyaminbass
Got it! Thank you! I've been downloading old models! @@sebastiankamph
great images but i never yet seen SDXL models show great human hands. Any chance you can make a point of demonstrating how to do this
Hands are still terrible unless super close up I'm afraid.
promt style is a patreon thing now.. seems like i cant download it for free
SORRY FOR being a unknowledgable european but can you tell me is there a local install for windiw 11 amd gpu to use this model on my pc ?!
ive tried to get into AI art few montsh ago but there wasnt any way to get it to work because i have amd gpu .....
I can't speak for using an AMD GPU, it may cause issues but the official Automatic1111 install file has been available as standalone install file. No need to download any dependencies and so on, it does it all in one go. Maybe you last tried before that became a feature? Might be worth trying again.
YT doesn't like links but it should be easy to find, or there's a link in the description for the one in the video, which is a good starter UI. 🙂
Can we say that a model is a holographic database of all images used for the training?
Not quite. Saying that it's a database of the images used for training implies that the images are stored in the model, but they aren't stored, the model merely learned from the images.
Unfortunately for me it doesn't work. I used the same prompts but all I get are deformed faces or not the same details
Any tips for us 8gb vram bros for trying to run XL models?
I modified the "webui-user.bat" and added the following arguments after set COMMANDLINE_ARGS: --xformers --medvram --no-half-vae. It helped with my img generation speed a lot (on rtx 2080ti, 11gb vram)
Comfy and Fooocus will run smoother for you then.
ok so updated a11 to 1.6 and using --medvram-sdxl 👍@@sebastiankamph
is there a 1.5 version of this?
My personal 1.5 favourite is epicrealism
@@sebastiankamph yea me too i get the most consistently beautifull pics with that model
mine just doesn't work at all :(
is it just me or does everything take ages with this model? or is this always the case for xl checkpoints? never used one before
XL models are more resource heavy, yes
So Ruined Fooocus is just like stable diffusion automatic 1111? We dont need it, right?
Yes, it's just another UI. You don't need it.
Is it just me, or does the model have a propensity to stretch the neck for tall images?
(With tall, I mean something like 1216x832, which is only 1.46:1, nothing extreme)
Many sdxl models do that. I use 1024x1024 if I want good proportions
What is RuinedFooocus?
It's a fork of popular UI Fooocus. See video description for link.
I need a better gpu, and start running SDXL.
Usually Civit-eye.
i fucking love your channel
Oh why thank you very much 🥰💫
For me its the opposite, been a photographer for 15 years and majority of models ive used give amazingly realistic results for portraits with the acception of occasional common body disfigurations with Ai.
I find the challenge is finding a model that can do creative artworks specifically to your needs.
A photograph i can be like it is what it is. But say like concept art, logo designs, character designs, etc. You need precise details to your vision. Otherwise its pointless to your project.
What does checkpoint mean? Can you use these commercially?
Checkpoint in this case is a Stable diffusion model. You can use it commercially, yes.
Thank you for your response.@@sebastiankamph
Had to delete it, with it as my checkpoint all my embeddings, and LORA's would disappear, and I did not have acess to them, when i went back to another checkpoint they came back, weird huh, anyway my favorite checkpoint is Lyriel, secound is dreamshaper
So much bleeding, green eyes and almost everything goes green 🙂
Per the CEO, it is Civi-Thai. :)
Sweet! 🌟
Make the page bigger.
HI SAB 1ST COMMENT? love your vids
First indeed! You're the real mvp 💫💥💫
@@sebastiankamph ممثاز
ممتاز
Juggernaut 6 sdxl !
Aghhh xl 😢
What I want is more amateur looking model. Like someone taking a picture of his friends etc. Most model looks way too professional.
Do we really believe they hand captioned 10,000 images? Wouldn't that be like a full time job for a small team over several weeks.
Combined with the fact they were also touting this as some kind of next level model on Reddit a few weeks back and then sponsoring UA-camrs makes me think they are planning on trying to convince people they have something special and then move future models to some kind of paid service.
So far I'm yet to see anything from it that I couldnt get from some of the other XL models.
I've spoken to the people that did the work and they indeed say it was hand captioned. I see no reason to make that up, but I can also not verify that it's true. It's a cool feature that improves the model, but not revolutionary. As far as I know, the models are not planned to go paid and from a business perspective more of a marketing tool.
@@sebastiankamph Ok that's good to hear. I'm still skeptical though as hand captioning 10,000 images is a huge and extremely tedious task and not something you would expect done for nothing in return. Even at one caption a minute which is generous that's over 160 hours of mind numbing work.
Oh, I don't think anyone is implying that it was done for nothing in return :) 💰@@justanothernobody7142
Woooo! RuinedFooocus!
The best! Just waiting for those gif features
Those millions of steps doesn't really show. Not much different than juggernaut. Actually worst in some cases for realistic results. The prompts shown here as examples are detailed in every model. Aliens in particular.
Not a best 😂😂😂 Nlight is far better
AIYou, I thought you've meant 3d models, which were made by humans, but no it's same hype of lazy stupid IA thing. IA will evolve human will degrade, not all, but many. Tools should help, not to replace human being at least in art.
TRY RuinedFooocus instead ;)
That's what I use 😂
@@sebastiankamph 🤒😵💫 Apparently .... I double checked now 😶🌫