Epic Realism Realistic AI Model In 9 Minutes – Stable Diffusion (Automatic1111)

so epic realism is another realistic checkpoint which can produce brilliant results with minimal prompting and we'll be looking at what it can do by running some tests to see how well it can handle our prompts drop a like to support the channel hit the Subscribe button and thanks for your support but enough talking let me give it to you bite-sized so epic realism is a checkpoint developed by epinikion which shows off some realistic looking images which do many things well including the eyes and mouth which look indistinguishable from the real thing the prompts in this checkpoint tend to be extremely simple which has the benefit of being extremely easy to use but often the disadvantage of not having much Variety in the outputs we can create this can be fixed with certain techniques so it's not the end of the world the description also has some advice on how to use the model courtesy of draw line and in a nutshell they recommend using simple prompts avoiding keywords like Masterpiece and keeping things overall simple with the embeddings and descriptions we also have a how to use guide which breaks down some of these suggested settings alongside including an upscaler which works well with this model and I'll have a link in the description they recommend two Samplers such as SD and DPM and these Samplers are how you go from a muddy image to a clean image each having their own pros and cons steps will also influence how clean your image will be with more steps resulting in a cleaner image and the amount you need depends on the sampler you're using but they suggest 20 the CFG scale determines how closely the image generation process aderes to your prompt and they suggest a value of five as anything higher can lose realism depending on your setup and lastly we have some useful extensions but there's not much to add other than I'm using the natural sin rc1 V version I started by replicating the example image to ensure we're getting the same quality and the results came out as we expected I'm using the version with the bait Vay and I won't be testing whether different vs have an impact as one has been recommended for for us the quality is great and I can't see any artifacts or quality problems in the image beside the background characters looking somewhat ghostly but with such a simple prompt it's expected as we're not controlling much in the scene going back to the suggestions within the description I wanted to test them out to see how significant they are starting with the recommendation to use Simple prompts and to avoid complex prompts as these can result in negative outcomes now it's debatable as to what a complex prompt is so I'm going to test two prompts starting with a short and simple one this results in a nice image as expected with nothing standing out as odd besides the comically long braids with somewhat merged with the finger using a more complex prompt the second image is also good but the braids are clipping through the jumper and through the braids beneath the ear it didn't seem to generate the background properly in that area as the colors seemed different to the cream bricks of the building I'm not sure whether my prompt was complex enough to trigger the negative effects but the checkpoint seems to handle prompts of a fair and reasonable length without any major quality impacts in the worst case scenario you could use this as a basis to generate an image and impt areas you don't like using another checkpoint or even using this as a refiner the next warning I noted was to avoid using any keywords like Masterpiece as they don't produce an appreciable change so I'll test a few of these keywords out using XYZ plots prompt Swap and replace function it was hard to tell whether the different prompts made any difference and I think it's safe to say that you're not missing out on anything by not specifying the quality of your images within your prompts this actually makes a lot of sense as you should get a high quality result by default without having to specify a type of quality but it can prevent this checkpoint from doing more stylized pieces and limits the experimentation of different prompts the page also said to add Asian and Chinese to the negative prompt if you're looking for ethnicities other than Asian but I didn't get any Asian looking images even when not specifying an ethnicity in my prompt but this did highlight another issue which is that all of the faces came out looking very syy and this isn't the end of the world as there are tricks to get different faces but the checkpoints ability to diversify a face can be important for avoiding the uncanny valley and this is something I noticed in their example images on their cotai where many of them had a similar eye nose and mouth and distinctly naso laal folds the last two warnings are in relation to using cinematic and one girl in your prompts as the former can remove natural effects while the latter can push images towards a render or anime style testing both they have a slight impact on the lighting and style but it's very faint so be aware that it will have some impact on the output but it's unlikely to do any damage or have any major impacts moving on to more standard tests I ran the CFG scale through some numbers where five is the recommended value and we've been using 6.5 from the example image I noticed that lower CFG values resulted in lower contrast less detail and a less saturated image while higher CFG scales did the opposite anything above eight started to look a bit off while four and Below had these odd patterns on the clothing with six being a nice balance of the best of everything therefore going with values between four and eight might be the best but this will of course depend on yourself and personal preference next we have the sampling steps and Samplers where I'll test both 2m cavis and SD cavas as the these are recommended 2m caras is said to give a more fantasy output while SD caras is recommended for more realism DPM Plus+ 2m Caris gives the best results in my view but requires a higher number of sampling steps for a good result compared to SD caras which has a legible image at 10 steps and even in some areas at Five Steps the clothing on the SD sillar looks somewhat confused while the 2m caras is more sensible and the backgrounds are a bit more interesting on the carass sampler so I'll likely use that one as my default option lastly looking at clip skip one is the best in terms of quality while two and three don't have any major issues but the background looks a bit odd and the clothing was all black and lacking any kind of variety despite us not specifying anything I usually tend to experiment with one or two clip skip depending on the results I'm getting and the checkpoint I'm using so it's worth experimenting with these to see what works best for your workflow now moving on to some other tests using celebrity names works really well if you want to diversify the faces being generated and this will likely be my go-to method for the most variety using something like prompt alternating to blend the prompts together although I'd highly recommend lowering the strength and mixing in more names so you can get a more distinct face which doesn't draw reference to any particular celebrity but rather looks like a unique individual on the art styles I struggle to get the art styles to pop through and had to use higher weighting which resulted in these very low quality images which seem to be fighting to maintain their realistic art style I wouldn't suggest using this checkpoint for a variety of art styles outside of realistic even if you're aiming for a realistic spin on anime cartoons or even watercolor on the skin tones I wanted to see whether the model could produce a variety of skin tones from light to dark and the model does an okay job but there's not a huge amount of variety across the different prompts you may be better specifying a person's name or using their ethnicity for more variation in skin tones but it does a good enough job for most purposes next I wanted to see how the checkpoint handles A random assortment of things like objects and animals it does a pretty good job on animals but struggles with vehicles and items which may include a character as that's the focus of this checkpoint you will likely need to put something like human or character into the negative prompt so that you get the objects alone and when it works it produces a great result although there may still be artifacts in some areas which will need to be imp painted out lastly I wanted to see how well this checkpoint can handle Landscapes and backgrounds on their own such as forests deserts and even space stations the results were fantastic with the Arctic and desert standing out as being almost completely believable with the variety in Plants shapes and how it handles the far away parts of the landscape the space station was also good but my example had Two Earths so you may want to prompt that out we also have an additional wing on the space station but it's a decent result which captures the bulk of what you would expect I'm also glad to see that there were no people in the landscape so I definitely recommend this for making Landscapes and background images but in conclusion I think this checkpoint is great for producing realistic results for pretty much anything but may struggle with unique faces which don't have similar features and generating images without people present but these issues can be resolve with some prompting and inating to remove what you don't want so give it a try and a big thanks to the supporters of the channel this is byze genius and I hope you enjoyed

Channel: Bitesized Genius

Views: 22,258

Keywords: epic realism, epicrealism, epic realism ai, epicrealism stable diffusion, epicrealism model, Epic realism tutorial, install epicrealism, epicrealism v5, epicrealism sdxl, epic realism model, epic realism v3, epic realism stable diffusion, stable diffusion tutorial, stable diffusion video, stable diffusion prompt guide, stable diffusion checkpoint models, stable diffusion checkpoint, stable diffusion checkpoint tutorial, stable diffusion checkpoints explained

Id: gILYw4spk8U

Length: 9min 10sec (550 seconds)

Published: Tue Jan 02 2024