the best REALISTIC models for Stable Diffusion

we're going over the best models for stable diffusion to create unbelievably lifelike images we'll jump straight in with one of my current favorites epic realism epic realism allows you to transform really simple prompts into stunningly lifelike results I find that this model really excels in capturing facial detail that other models simply brush over so here on the download page we can see what kind of images the users have generated remember you can click on any of these images to see what prompts other users have used to create theirs all right let's move into automatic 1111 to see what we can make firsthand initially strike for Simplicity and avoid adding extra keywords into your prompts for example words like Masterpiece best quality and 8K shouldn't really be included as they add no noticeable difference to the outcome however words such as cartoon painting and illustration should be included in the negatives as they can detract from the realistic qualities the key to maintaining the perfect balance between quality and realism lies in fine-tuning several parameters I keep the steps above 20 and sometimes I even go for higher numbers like 14 and 60. particularly if I'm encountering image errors or artifacts like malformed limbs or blurry faces in regards to the CFG scale the author recommends you set it to five as cranking up might compromise the realistic feel depending on your prompt The Chosen sampler and the number of steps of course the sampler is your playground any sampler works but for the extra dose of realism I find that using DPM sde Caris or dpm2m Keras works the best other Samplers like DPM fast work well enough as well I'd recommend you experiment and see for yourself as depending on your prompt each sampler could outshine the other to achieve the best resolution I'd recommend using the high res upscaler I found success using either the nmkd super scale or nmkd faces each with a denoising setting of 0.35 and an upscale factor of 2. these two upscale has simply improve the level of detail on the generated image and the denoising strength is how close to the pre-up scaled image the final output will be so for example I generate an image without an upscaler it's a fairly good image but the face looks a little bit smudged for my standards especially the lip it's actually because the image wasn't a high enough resolution to give the face good detail so let's make it look better with an upscaler I'm going to set it to a denoising strength of 0.35 and an upscale factor of 2. and we can see from the comparison that the face of upscaled has way finer details that the original just doesn't have so yeah I'd really recommend using an upscaler I've linked both the ones I just talked about in the description and I recommend you download them and experiment with both of them here are some additional tips I've learned to truly harness the power of effort realism you really need to make effective use of your negatives this not only helps to add the realism to the image but also helps to Define what you don't want to know your image as we probably all noticed almost all the realistic models in stable diffusion tend to be biased towards creating East Asian women so from aiming for an ethnicity other than Asian I simply add Asian comma Chinese to my negative Light Shadows and other intricate details are captured excellently by the model without any extra effort so there isn't any real need to add lighting keywords like hard light or cinematic lighting and for a more natural effect refrain from using the term cinematic at all in your prompt and I've also learned that over describing the face often yields less desirable results and lastly I'd recommend you use the Epic realism help allora by the same author it just further helps over realism to create extremely lifelike images to download epic realism or any other model for that matter simply navigate to the link I put in the description once you're here click download and let it download into your stable diffusion web UI slash models stable diffusion folder once it's done open automatically level 11 and just select it at the top epic realism is my favorite model at the moment I'm sure you can see why it produces mind-blowing results very easily however there are still a lot of extremely powerful models so moving on to our next Contender let's look at the Magic Mix model while not necessarily topping the charts of my personal favorites it undeniably has its own unique strengths Magic Mix truly sells in the realm of dramatic and dark lit scenes really bringing out the moodiness and mystery of your Generations however it's important to note the model's limitations especially when it comes to facial generation with that proper prompting Magic Mix will almost exclusively generate East Asian women and most of the time the model tends to lean toward a uniform and unrealistic Tick-Tock slim face filter look however if this happens to be your preferred style magic Mitch might just be the ideal model for you here's how I found to optimize the use of Magic Mix with this model your options for a sampler are Eula a Euler dpm2m Karis or dpmsc cares I find that all of them work fairly well and can produce great results when it comes to the number of steps I found that the sweet spot is between 20 and 40. to make your images look even better for upscaling I'd recommend either using the nmkd faces or nmkd super scale I found that setting the high resup scale to 2 and the high-res steps to 15 works best with the denoising strength I find that anywhere between 0.1 and 0.5 produces good results so now I generate some example images using major mix I'll also generate their non-up scaled versions so you can clearly see the difference so this is an image I just generated without an upscaler and now this is the same image with the upscaler applied really nice thus far but let's continue looking at some of the other settings to further improve our Generations the convex shell is another important parameter to adjust I find that a range between 6 and 8 typically yields the best results for your positive prompts the author actually recommends using terms such as best quality Masterpiece and photorealistic as they actually do make a difference with magic mix with the majority of other realistic models terms like that really make any sort of meaningful difference for the nectar prompts which help to shape what you don't want to see on your image I find that the textual inversions NG deep negative and bad hand V4 improve the images deep negative basically just lessens the chance of your image going totally insane with malformed anatomy and really helps the model from looking too cartoony and bad hand simply improves the hands which stable the fusion and frankly all air tools still struggle with everything I'm talking about today is linked Down Below in the description although magic mitts may have its quirks it's a fantastic one of creating images with striking lighting effects and Atmospheric settings just keep in mind this inclination towards a specific facial style and you can certainly craft some great AI generated artwork I'd just like to say I'd massively appreciate it if you liked the video if you found value in it thus far anyways let's now turn our attention to a model that thrives on versatility and dynamicism with analog Madness this model really breaks away from the crowd with an ability to generate images of ordinary individuals a rather refreshing alternative to the supermodel Renditions often produced by the other popular models including those talked about in this video The Power of analog Madness lies in the potency of the prompts provided the more Vivid and robust your prompts the more captivating the output becomes let's break down my exact workflow with this model I find that the sde Cara sampler proves to be the ideal Choice when working with analog Madness and for an optimal balance between details and computational load I maintain a range between 25 and 35 steps when it comes to the conflict scale I find that the default setting of 7 offers the best results and terms of realism now crafting the perfect prompts is crucial for the nectar prompt I've observed that keywords such as 3D Max grotesque and desaturated work well to make the image more realistic in terms of color and just general composition alright now let's look at what this model is truly capable of let's consider a positive prompt like this hyper realistic GoPro action photo of a 20 year old Dutch woman with black hair looking at camera which have adjusted the weight of using brackets wearing a leather jacket on the moon and the lock star I'm keeping it extremely specific and pointed prompt stars like this work extremely well specifically for analog Madness but not that well for the other models I've tried including epic realism and Magic Mix that's a pretty good result now it certainly didn't get it perfect as you can see the arm goes slightly strange however it's a good first result to demonstrate Anna love madness's strength in generating realistic non-modalesque figures while still maintaining an incredible level of detail and complexity analog Madness with its refreshing take on AI image generation truly broadens the horizons of what's possible in this realm by playing with your prompts and keeping the steps in Conflict scale within my recommended parameters you can certainly tap into its potential to create a wide variety of realistic and unique images and that's all for today check out my website for more guides and thanks for watching [Music]

Channel: James Beltman

Views: 38,031

Keywords: ai art, stable diffusion, stable diffusion tutorial, ai art generator, ai, midjourney, ai art tools, artificial intelligence art

Id: hYihkH895GQ

Length: 8min 44sec (524 seconds)

Published: Wed Jul 26 2023