You can use my model directly at : Tensort.Art | Mage.Space
_____
On top of that, SDXL is only in its early stages, I may have many shortcomings. I hope you can give me suggestions for corrections in the next updates. If you find it useful, please vote 5 stars and post an image to help me, it's really important, I thank you all very much
_____
📖 Checkpoint Info
Info : This is a training model based on the best quality photos created from SDVN3-RealArt model. With the desire to bring the beauty of SD1.5 to SDXL model. I added a bit of real life and skin detailing to improve facial detail. Currently I have two versions Beautyface and Slimface.Detailface version is an upgrade of Beautyface, it creates more skin details and creates better realism than the beautyface version. With the Beautyface version - the model face will have more diverse angular lines. In contrast, the Slimface version will have a slimmer and easier-to-see model face, suitable for Asian faces.
_____
You can refer to some of the indicators below to achieve the best image quality :
Steps : > 50
CFG : 9-10
Size : 768x1162 px ( or 800x1200px )
You can also use hiresfix ( hiresfix is not really good at SDXL, if you use it please consider denoising streng 0.2-0.3 ) or After Detailer
Sampling : Euler a or DPM ++ SDE Karass
📝 Quality Prompt :
Base quality prompt :
(masterpiece,best quality, ultra realistic,32k,RAW photo,detail skin, 8k uhd, dslr,high quality, film grain:1.5),📝 Negative Prompt :
(worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), tooth, open mouth,💡 Concept reference:
Portrait Lady Real :
(masterpiece,best quality, ultra realistic,32k,RAW photo,detail skin, 8k uhd, dslr,high quality, film grain:1.5),
{short hair | long hair}, __*/haircolor__ hair,lady, brown eyes, __*/colors__ __*/Wear/*__, freckles,
(__*/colors__ gradient background, simple background:1.3)_____
Also you can test those prompt tags quickly with my model here : Tensor.art
I have written a colab site that integrates all tools for you to use stablediffusion without configuring your computer, you can refer to : Colab SDVN
I have provided an album with more than 2000 original photos with full prompt, you can refer to understand how to use SDXL : SDVN-SDXLPrompt Kit
In addition, some of my models are available on the Mage.space platform, you can refer to: SDVN Mage
_____
On top of that, SDXL is only in its early stages, I may have many shortcomings. I hope you can give me suggestions for corrections in the next updates. If you find it useful, please vote 5 stars and post an image to help me, it's really important, I thank you all very much
_____
👨🏻💻 Creator : stablediffusion.vn
📢 Discord : Discord SDVN
📢 Facebook : Group SDVN
☕️ Donate me : Bank | Buymeacoffee
Description
In this version, I added a little element to give the model skin more detail. The model face lines in this version will be a bit more diverse and angular than the slim face version
FAQ
Comments (46)
Still totally airbrushed results. Is this intentional or the basic model is just not allowing proper texturing and look?
Maybe it needs more rounds of training and teaching about materials. At the moment, the tools to support train and tweak are quite poor so it can be balanced between the good factors. I could have made the material better, but the model face is pretty deformed :( . If there were fine-tuning tools like a 1.5, I would probably do better.
I guess the problem is that there is still no trained refiner. The only awesome results superior to SD 1.5 I can generate with base xl and xl refiner.
SDXL makes perfectly non-airbrushed results with simple language prompts but it seems everyone keeps trying 1.5 prompts on SDXL expecting the same results when it was trained completely differently.
@VolatileSupernova I get it, I'm trying to optimize the prompt, but trying to be more concise will also affect the results more or less, creating the rate of unwanted images.
Part of that is because to train SDXL needs much higher res training images, so people have to re-optimise their datasets. Good skin texture will come, don't worry 😁
@morbuto I have optimized an even better quality, along with an improvement in the generation stage. Please look forward to my upcoming updates
@VolatileSupernova do you have any example for simple language for non-airbrushed face of SDXL photography? for two weeks I got that super smooth photoshoped like faces.
THY, i will take a look...
but these "grain:1.5" does not work, try it youself ;)
Yes, maybe some keywords don't work, I'm optimizing more and more to update the tutorial
@StableDiffusionVN i mean that number its seams not supported in SDXL ;)
@sevenof9247 no, it really works, I tested it
@StableDiffusionVN okay, i tested it with original base model, 1 or 1.5 or 2 is no different, also with freckles or other words ... hmmmm
@sevenof9247 The way it works is definitely different from 1.5, but I still find it impactful, at least the freckles seem quite different.
@StableDiffusionVN i meant the numbers 1, 1.5 nad 2 also masterpiece and so on, hmmmm why no one tell us how we can feed SDXL ? :D
DPM, not DMP!
Yes, I fixed it, thanks for the feedback ^^
why that ugly moles on faces again?
I think that would be more realistic :D
and they are cute :)
There is a negative prompt if mole is not your thing
They are more realistic. They used to be called "beauty marks"! Girls used to draw them on with eyeliner if they didn't have them naturally. I can't believe how some beauty standards have flipped 180 degrees. Now many girls IRL slather on thick plastic makeup and try to make "shiny" skin. Draw sharp triangles for eyeliner. Even their eyebrows look like they were drawn on with a magic marker. 😉 So yeah, even the smooth faces here like modern photos.
Did you make those arts with ComfyUi? Cuz it's not working on a1111 for me
I am creating a lot of images in comfyUI, it's very beautiful when created with comfy, some of the images below are created with comfyUI, you can refer. In the upcoming update, my review images will mostly use comfyUI
@StableDiffusionVN oh i see. did u use a1111 too?
@linova4375539 i'm using both
With reference to your reference picture prompt.
Unless you are looking to lower the quality of the image with film grain intentionally... you are wasting resources by having it, and weighting it. I've never seen any image actually produce an example of the grain effect similar to what google shows what film grain is.
Thank you for your feedback, I will optimize the prompt in the near future
Yeah I don't think there are a lot of good modern references for this. If you haven't had to push process Tri-X in HC110B, you don't know what obnoxious film grain is, and why you really don't want it. 😉 That's (film grain:1.5) at least. 🤣 Not that many people shoot film today and digital has it's own artifacts. None I'd want to reproduce.
I look at the pics below, don't feel like anything special about SDXL yet.
If you pay close attention, it will have a better story in small details than SD1.5 . Small details like lips, eyes, nose are better represented. In this version, the skin quality of the model is still not very good, but I have improved it in the next update. Furthermore, SDXL can be very creative, the pictures below are just a small part of what SDXL can do.
I'm looking forward
It will take months lol of course it won't be better now.
Have you seen the base model 1.5 when it came out and what the community did with it?
Give it a couple months. SDXL is better in every ways, especially more flexible in prompting. We just need the community to do its thing
The feeling of plastic is particularly strong
@weapon98827 I will try to fix it gradually
It's nothing special despite being much less mature. It means it's a great starting point. Come back in a couple of weeks.
@whatever_crap706 thank you , i will update more soon , train SDXL is really expensive
Many practitioners argue that Stable Diffusion is incapable of facilitating genuine commercial creativity, primarily due to the substantial errors present in the images generated by Stable Diffusion 1.5 version. These errors encompass not only inaccuracies in finger and body structures, but also involve intricate interlocking formations and logical relationships of light and shadow. The excessive amount of detail further renders these errors unnoticed by many non-professionals.
Screw what others are doing. Follow your own path. I don't do girls either. I just ran it for the first time and plain vanilla SDXL output of one of my old architecture prompts looks like a good 1.5 finetune... but... 1k x 1k no HRF and it's super fast. And... I got ONE HOUSE. Not an apartment complex. 🤣. That was usually 1 in about 30 with HRF. The refiner adds details and it's pretty fast too. So now I'm finally here scooping up all the XL finetunes and Lora I can find.
@sh1201 so commercial creativity requires fingers? If you mean merely duplicating what a camera can do, sure. But I've had a camera since I was a kid. I don't need another one. I don't waste time trying to create perfect, normal, photorealistic portraits of pretty young girls. Sure you're going to see "errors" there. I'm a commercial 2D and 3D artist. I have a general contract with a client that lets me entertain them with any tools I see fit. They're delighted with what SD can do. I have work up on stage on Peter Gabriel's current concert tour. If by "professional", you mean porn or pinups or patreon copy some other artist but tack on huge boobs, then ignore me, that's not my thing. None of my clients have ever wanted anything to do with that world.
I cant get this to download in my colab for training. What gives ?
@sh1201 I guess many of the SDXL shortcommings can be easely corrected with the help of photoshop beta new features like generative ai and remove tool. It really helps get rid of those small imperfections, if you really intenting to use your images commercially.
@parallelepipedon I never mentioned fingers, nor is that what I intended to convey. Please refrain from guessing and responding based on assumptions. The main issue with SD1.5 lies in its poor structure and lighting. This problem has been significantly improved in the SDXL version.
@sh1201 psst.. quote from your post earlier: "These errors encompass not only inaccuracies in finger and body structures, but also involve intricate interlocking formations and logical relationships of light and shadow."
I think you might've mentioned fingers there ;-) but it's all good. <3
My personal experience is that I'm way more controlling and getting better finesse in the art I make compared to what I'm trying to get done in commercial versions. Using a mix of ComfyUI and Automatic1111 and eventually finetuning (sometimes not even needed) in Photoshop is a blessing (and I'm not even using photoshop's generative fill).
It's a good model to use. Cheers! <3
@nicolaswuffwuff My mistake, my point is that fingers are not the main problem of sd1.5, and fingers can be easily fixed comparing to other mistakes.
Combining Comfy and A1111 is also my method for generating images for personal enjoyment. I believe this could represent the future of visual content creation, especially with the recent introduction of the StableSwarmUI project by Stability-AI.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.