Hey, you can get a sharper picture and some other optimizations with the V2 version! However, the V1 version is more creative, but the screen clarity will not be as good as V2 as well as redundant generation in some specific cases.
Although SD3 has been released, there are still considerable limitations and uncertainties. XL is good enough now and has some potential for improvement.
In the making of this model, I wanted the new model to come close to the excellent detailing of SD3, and to outperform the current SD3 in terms of aesthetics and structure.
My goal is to have the model be one of the best XL Reality models out there. Though that's not very realistic haha.
Features
Specializes in realistic photo-like image generation
The generated images have good structure and exquisite detail performance
The color performance and contrast performance are excellent. The semantic understanding ability may be a little better than most XL models.
It is not good at non-realistic painting styles such as illustrations and animations.
High dynamic range image generation from pure black to pure white can be achieved without Model Sampling Continuous EDM.
The comprehensive generation can be relatively balanced, and various types of real images can be completed well
Whether it is simple or complex prompts, images can be generated well
Test parameters
CFG range: 1-8 ranges have achieved good generation quality. It is recommended to use 2-4 ranges. The generated images will be more realistic and not too strong.
Sampler: Compatible with most samplers
Scheduler: Compatible with most schedulers
Resolution range: It can support high-quality generation of any image within 512 to 1536 resolution, and can achieve up to 4096 resolution image generation with HiDiffusion.
steps: It is recommended to use more than 20 steps to achieve better quality without using accelerated lora
Lora: Only accelerated lora has been tested, which is perfectly compatible. Other types of lora have not been tested. It is estimated that the compatibility with illustration anime style lora is not good.
The scheduler used in my test is: AlignYourStepsScheduler, and the sampler is: dpmpp.2m_sde_gpu. It is only for reference. This does not mean that other samplers or schedulers are inappropriate. Their effects are also very good. You can choose according to your preferences.
Since Civitai cannot automatically identify the generation parameters from the workflow images, you need to add them manually one by one. Please wait for good news. Or you can go to the Google and OpenAI websites to find the corresponding prompts. Most of the prompts for the example images of this model come from them.
You can also view the used workflow within comfyui by downloading the original image.
Please note:
The main goal of this model is to achieve strong generic generation capabilities. Therefore, it will not be trained specifically for a single concept, for example, to enhance the results for portraits, as this would destabilize the model for other rich concepts to a certain extent.
However, this does not mean that it will be bad at generating a specific concept, on the contrary, it will still be able to generate good images and have more creative freedom than a specialized model.
If you need strong control over a specific concept you can use the lora technique with this model, just give lora a small weight and it can be sensitively influenced to change the model from a generic generator to a single concept-enhanced generator.
Recommend a set of generic cue words that I use regularly:
Positive:
focus, masterpiece, very clear, best quality, High quality, high detail, high resolution, error-free, Cinematic compositionNegative:
NSFW, deformity, low quality, low resolution, blurry, unclear, wrong, watermarked, noisy, soft, deformed, ugly, deformed, mutated, Lens shake, distortionIt's a gift. Have fun.
I'm glad you've made it this far. Let me clarify that this model was customized for my own needs and is not designed to cater to everyone. It is solely intended for the generation or re-creation of photorealistic images and has not been specifically optimized for any independent concept such as NSFW content. It is not suited for a wide variety of artistic styles; it focuses only on realistic imagery. Additionally, there are plenty of other excellent models within the community that are also used for generating realistic images.
Given the limitations of the XL architecture, it's rare for one outstanding model to completely outperform others, except for some truly poor models. Most high-quality models have their own strengths and weaknesses.
So, please assess whether you truly need this model before downloading it. If you already have another excellent realistic model, consider whether you really need to download another one. I don't want you to waste your valuable time downloading and comparing models, only to then complain about being tired of new model releases without seeing significant improvements in results. If you enjoy exploring out of sheer interest and don't mind how many models you download or how much time you spend, then by all means, feel free to use it.
Description
Ability to generate better realistic details of the subject (it is now possible to generate minuscule details such as dust on the surface of the object or even tiny hairs on the skin)
More natural looking sharpening
Reduced the probability of duplication of the subject object when generating images with a very wide horizontal orientation.
(It is really difficult to eliminate this phenomenon completely, at the cost of reducing the richness of the elements that appear in the image).
More natural light and shadow effects
FAQ
Comments (9)
extremely impressive! prompt adherence is great, i dont find myself fluffing up the prompt to get what i want
as a recommendation, i would love to see very different diversities if that makes sense, i find that reguardless of if i say "korean, japanese" etc for example, i mostly get images of hong kong type locations
this model is super promising overall, love it!
Hey there, really appreciate your thoughts and feedback on the model! It's great to know it's making an impression.
I've done some testing, and I haven't noticed the diversity issue you mentioned, but I'm definitely open to the possibility that it could be there. I'll keep an eye on it and do some more probing.
If you have any specific instances or prompts you can share, that would be awesome. It could help me understand the situation better.
Thanks again for reaching out and for your support. It's this kind of input that helps push the project forward!
@LragonStarr i misspoke in my previous message, what i noticed was with prompts like "masterpiece, very clear, best quality,high detail, high resolution, error-free, a beautuful chinese woman, portrait", i could replace the "chinese" with previously mentioned cultures, and found that the images were pretty similar, giving very japanese inspired portraits
further more, adding "city background, gave very tokyo inspired images reguardless.
its very much not an issue, just something i noticed with simpler prompts and very specific cases, adding detail with, "new york" for example works great, and produces stunning images as usual!
I dono man im gettin tired of people calling merges "the next best SDXL model" only to use it and feel like im just using the same models as before.
Hey, I understand that the hype of "the next best thing" can be tiresome, especially when it doesn't quite work. As a solo creator, I'm equally challenged to innovate as much as technology allows.
I put a lot of effort into building a model that meets expectations as well as possible, utilizing the latest relevant technology and fine-grained control of each layer of weights and or CLIP, and even training to produce tangible improvements, even if they are minimal.
It's hard when models don't work as well as they should, and I know that not all models work as expected. That's why I'm cautiously optimistic about my product - it's well-built, I want it to be perfect, but it may not be, and I can't help but want to share my work whenever it's effective and produces tangible improvements, and will do the same in the future when the work I'm building on for the next generation pays off.
Nothing in the world is one-size-fits-all, keep an open mind and enjoy it all bro.
while I have to compare results to v1, this model now looks like all the other models. Not sure what you merged It with, but by adding whatever it is you added, this looks the same, and there is less creativity now.
Regarding the issue of the V2 model's creativity in terms of visual elements, as I mentioned in the update notes, this was a compromise made to address another issue. As for the so-called similarity of results with other models that you mentioned, I don't understand. I have conducted extensive comparative tests with multiple mainstream models. While I cannot guarantee that the results will always be better, in most cases, this model slightly outperforms the others. The approximate ratio is that for every batch of twenty images, this model achieves better results in more than ten, about the same in five or six, and only a few are less impressive. Whether it's the ability to understand prompt words sensitively, avoiding the generation of incorrect structures, or the color atmosphere of the image, all have been tested and verified. Of course, there are many models in the community, and I cannot compare with every single one. I have only conducted experiments with some popular models, such as realvisxlV40, juggernautXL, sdxlUnstableDiffusers, leosamsHelloworldXL, etc. Compared to their results, I am more satisfied with V2's performance in generating realistic images. They are also excellent models with their own outstanding abilities, such as adaptability to NSFW themes, the ability to adapt to a rich variety of image styles, or the exceptional results in generating portraits. My original intention for creating this model was to meet my own demand for more advanced realistic images. I believe that taking matters into my own hands can more precisely control everything according to my needs, rather than downloading a large number of models one by one for testing. I just happen to share it with everyone for use. I neither force anyone to use it nor generate any income from it. Some people may be tired of the endless stream of new models, while others maintain an open mind and continue to explore new things. It's great if someone likes it; if not, they can simply ignore it. After all, nothing can satisfy everyone.
@LragonStarr I have 164 tests, I compare it to each model. After a while you get to know what you'll get a feeling of what it normally looks like. The first one is totally different looking, its creative and fun to see what it makes. But this one undoes that. Now I get the typical results that most models give. Which is disappointing. I'm not sure what you added to this one but now it just has that "looks like all the other models" look.
@frankmike So what you’re implying is that you hope the new model can maintain the results of the v1 version while making adjustments? If that’s the case, then I would say, that should be a V1.1 version. I did create an early version like this during the preliminary stages of the V2 upgrade work. It had the creativity of v1 while eliminating the blur of V1. However, I still needed to address another frequently occurring issue of object repetition, so I abandoned this version.