Cyber Realistic Realistic AI Model In 7 Minutes – Stable Diffusion (Automatic1111)

cyber realistic is another Contender for a fantastic realistic style checkpoint and in this video we'll be running some tests and seeing whether we can achieve some fantastic results be sure to drop a like on the video and subscribe to support the Channel's content and check out the patreon but enough talk let me give it to you bite-sized so cyber realistic it's a realistic style checkpoint developed by cyberia and Promises a versatile photo realistic model which was regularly tested by blending vious models to achieve the desired output the checkpoint provides us with a range of sample photos which look good and distinct from one another including some food items environment pieces and a very good boy one of the key strengths is in the ability to effectively process textual inversions and luras providing accurate and detailed outputs we also require minimal prompts to get good results and the page ends with some suggested resources including the Cyber realistic negative embedding which we will need to download the description was brief and to the point but didn't provide any recommended settings so we'll have to base our tests on the settings used within the provided example images moving on to our first test I copied over the generation data from an example image to see whether the checkpoint was functioning as expected by giving us the same result as the auor and our image came out very similar with a slight difference in the posture and Hue that has lost some of that yellow glow but nothing world ending and the quality is still good I also wanted to see what quality differences we may get if we remove the Cyber realistic negative embedding and it does make a significant difference where we have improved lighting and some error connections with the consistency and multiple candles so worth using the Cyber realistic embedding with this checkpoint and it may even provide good results on other checkpoints now let's move on to testing different settings so we can see what compromises we can make between performance quality and getting optimal results I've generated my own image which had no quality problems and maintained the excellent results weeded so far so my confidence is high on the quality front starting with the sampling steps I tested a range from 10 to 50 which also leaves some room to compare the resulting images and upon comparison there's no noticeable difference beyond the 20 steps the images all have the same degree of quality with minor differences here and there but nothing significant enough to Warrant a higher sampling step and in some instances a higher sampling step produced the worse result on the necklace but at the same time a sampling step of 10 produced a far lower quality result especially with the glasses which had this orange artifact in the lens and the clothes were all jumbled together so worth going for around 20 as the optimal result now turning to Samplers I've tested our regular options which include the two popular DPM 2m and SD cars samplers and the Ula a and DD IM Samplers to see how they perform with this checkpoint the best result came from 2m Caris as a checkpoint recommended but DD IM also provided excellent results the images look very similar and capture what we'd expect from our prompt but SD caras added a denim jacket while Ula a gave us a plain background I think to M Caris and DD IM are the clear winners of this test looking at the CFG scale we get decent results across the board but I think 5 to9 gives the best results without any harshness or losing too much detail I think lower values work far better than higher values then lastly for the clip skip one gives the best results across the board while two and three look odd with the mouths lighting and composition four went and did its own thing where the top image lacks the glasses and the bottom isn't too bad but I think the clip SC of one is the best option for this model moving on to some further t test running a few prompts for the skin tone provided us with some nice distinctions between pale to Black with purple not working to change any skin colors I do wish we had a darker skin tone as black is more of a brown but there is a nice selection here to choose from regardless with a faint difference between Olive and tand we also didn't get any other changes which is good as we only wanted to modify the skin tone rather than the clothing or background but on the topic of Darker skin tones I ran some ethnicity prompts and we did get a slightly darker skin tone when using African or Jamaican than our prompt our other prompts like Chinese and Japanese looked accurate alongside Indian but the Latino prompt look like a tanned version of the other faces are not too distinctly different so perhaps someone can pick up on What's Missing but it seems to do a decent enough job of understanding how people in generalized areas may look on the age prompts these work completely fine and gave us a clear distinct set of looks depending on the age we specified with younger mid 20 not having much of a difference in terms of looks and on our style prompts as you probably expected we aren't getting any Unique Kind of styling happening with this checkpoint which gears itself more towards purely photographic SL realistic Styles likely benefiting from prompts like film grain and dramatic over anime and watercolor testing objects we get a range of interesting results across three different seeds starting with the water we get a normal looking glass of water a nice cold looking glass of water and then a type of water Logan PA would sell to Children two of the pencils came fresh out of a detective Thriller while the middle one looks more like a crayon and now chewing gum looks more like Jaw Breakers or those gumballs you'll find in those public vending machines from the olden days the results aren't bad quality-wise but there is a diverse interpretation of what I would have thought would be simple objects which may be a positive or negative depending on your objectives for this model then looking at animals we get some high quality results which do vary in terms of accuracy the rabbit looks fine and other than the long neck I don't see any major problems and the snake is okay but there are some artifacts and it does seem to get confused with the anatomy then with the Scorpion we ended up with some kind of locus grasshopper type insect which looks nothing like a scorpion so good quality in terms of the image back accurac is a problem from this test our lastest test will'll be trying out some Landscapes and the results were pretty good in all honesty the beach and forest look great and simple but the Super Market caught me by surprise as we have the interior with stock produce while not being detailed does look convincing at a distance we've got lots of variety including fruits vegetables packaged Goods signs lighting ventilation and Reflections not a bad set of results I think in summary this checkpoint produces good results especially with people but does lack accuracy in some other areas but hopefully you found this overview helpful and if you did then drop a like on the video subscribe and check out the patreon for image files from all my videos this is by siiz genius and I hope you enjoyed

Channel: Bitesized Genius

Views: 3,241

Keywords: cyber realistic, cyber realistic stable diffusion, cyberrealistic, cyber realsitic, cyberrealistic negative, cyberrealistic stable diffusion colab, Cyber realsitic stable diffusion, cyberrealistic stable diffusion, stable diffusion, stable diffusion checkpoint models, stable diffusion checkpoint download, realistic model stable diffusion

Id: ha4fI_MoD6I

Length: 7min 14sec (434 seconds)

Published: Thu Feb 29 2024