Unstable Diffusion Checkpoint In 11 Minutes - Stable Diffusion Tutorial (Automatic1111)

unstable diffusion describes itself as an extraordinary model which brints together boundless creativity and unpredictability of unstable diffuses with enhanced versatility and styleon loocking capabilities of the Proteus a very mystical and encouraging description so let's see what this very popular model can do in practice by giving it a spin so you can decide whether it's right for you but like the video and let me give it to you by it size so sdxl unstable diffuser was made by yam and we're greeted with an introduction to the model we'll be testing Nill mania this is a blend of two models being the unstable diffusers v11 and data voids Proteus run diffusion checkpoint which I have to check out another time now one of the strengths of the model being advertised is the clip architecture from the Proteus model which enabled the model to generate art ranging from anime to photo realistic and this is demonstrated through the range of sample images which we'll get into later but taking a quick glance you can see the range of possibilities with this checkpoint there's an update Note for the Nill Mania version 12 which summarizes that the model has improved its handling of the CFG scaling with a range of 5 to 50 and that pony diffusers trigger words can be used but these are experimental Pony diffusion seems to be a popular checkpoint for safe and non-safe for work images primarily focused on furry and humanoid Concepts using simple natural language prompts for the sake of my YouTube channel I will not Venture any further into this model but the key takeaway is that it's wellmade and stable diffusers is borrowing some of that capability there's some information on how users should approach the model encouraging experimentation and the author ends the update Note by jerking off the audience to an unforgettable Journey Through the Realms of imagination which I'm looking forward to greatly now there's even more information regarding what an unstable diffuser is before the description vent into explaining each of the model versions and most of the descriptions can be summarized as it's versatile download it I'll skip down further and you can see the Creator's note and the list of reasons why the checkpoint is better than sdxl 1.0 and if we skip down further we can find some instructions on how best to use the model which is what I was looking for so we can get into some action we have a la called Excel yamama style which comes recommended for anyone looking to reinforce the style of the checkpoint or another a checkpoint of your choice restore face is not needed to generate high quality images and the checkpoint works well with Eclipse G of Evil 1 or two we also don't need to use any refiners when using this model so we should get good results on its own there's also a couple of quotes from y m Grimlock inserted alongside the author's personal ratings eventually we get to the thing I will searching for to begin with the recommended settings to generate the image and I feel like I've climbed the mounting of exposition to get to this point although my excitement levels are high and my expectations higher the auor recommends that we use a resolution of 1024x 1024 or 16x9 4x3 or 6x13 aspect ratios you can find the aspect ratio calculator online if you want to know what those mean in terms of width and height the steps are 35 to 150 with a warning that images with steps below 30 May produce artifacts or Weir saturation and the up scaler of choice is full Hardy or 4X Ultra sharp and I have a link to those UPS scalers in the description box below those UPS scalers should be downloaded and dropped into your web UI installation folder models ESR folder as a pth file and the high res up scale is limited by your GPU but 2.5 is a good figure also don't forget to download the sdx LV which will need to go into the web UI installation folders model V folder ens sh you have it selected in web UI vas settings so you don't produce any artifacts in your images if you're looking to make naughty images then use an sdxl laa as this model isn't focused on that type of content but it is capable of doing it and the description finally ends with some V10 turbo settings for those using the V10 turbo Edition version of the checkpoint which is available for download but now that we've finally covered all of that let's dive into some testing to see what this checkpoint is capable of achieving my my first test was to generate the example image to ensure we're getting the same results and everything is functioning correctly I've opted for this magical cat which uses a clip skip of one and taking a look at our prompts we have a mixed bunch of words being used to make up this whole image we've got those enhancers like Masterpiece while describing the style of the image including the colors and even an artist to draw inspiration from we also have some extensive but understandable negative prompts looking at our image up close it looks pretty accurate to the sample image provided by the author and from a quality perspective I can't see anything wrong with the image in terms of artifacts or errors in the anatomy now we'll move on to testing how well this checkpoint can handle different art styles using the recommended settings from the checkpoints description but first let's test out how the different settings impact the quality of the image starting with the steps as anything below 30 can cause artifacts I'm going to test 10 to 50 to see how bad the quality is impacted now It's tricky to notice any artifacts in these images but there is a notable quality increase from steps 30 where we're getting more details in the cat's face and less blending between the fur and the clouds 10 steps looks fine but lower quality and 20 steps does have some of that cat's fur missing on the right hand side but you might struggle to notice It ultimately you could get away with less than 30 steps but I go with 30 and above since the quality is much better then looking at the CFG scale I noticed that we get get a higher contrast and somewhat harsher results with a higher CFG scale and more softer results on a lower CFG scale so worth playing around with its value to get what it is you like next we'll look at the Samplers and I've chosen two DPM Samplers alongside Ula a and DD IM four popular options in total so we can see a comparison I think the results across the images came out well and we have a lot of diversity from the cat's face to halfer body a more abstract piece and a sharper image I think two M Caris and SD caras take the lead for being the best looking while Ula a has some other objects like flowers and a butterfly and DD IM is the same as 2m caras but a bit sharper which could be added manually then lastly looking at the clip skip we get pretty similar results across the four clip skips so not much to say as it doesn't seem to make much of a difference so next I'm going to be using XYZ plot to test out a few Styles but I've hidden the legend so we can see the images in better detail although it should be obvious what art styles are from left to right being anime realistic 2D and pencil now getting the art styles to pop did take some experimentation and explaining the image in different ways mainly ensuring the art style was mentioned multiple times in the prompt in different ways such as using both pencil art style and graphite to push it further but when you do get the results popping through it does a pretty good job although the prompt wasn't interpreted exactly but I did choose a very abstract prompt to use of plants dancing with smiles two of our images have a woman in the field while the other two are close-up shots of the sunflowers in different styles and I think the last two are my favorites this was created using a short set of prompts and I did notice that when changing the seed the accuracy of the images did change so you may need to test out a few seeds or give some of your prompts a higher waiting so you can get a stronger effect that's more reliable across seeds but ultimately this checkpoint can handle a range of styles well and will likely work wonders for in painting images due to its versatility now I wanted to check out the skin tones since we're exploring the results we can achieve with this model and I lowered the upscaling value as a generation time was longer than I liked we did get a fair distribution of skin tones across our test image and two seeds with pale and white looking quite similar but the others having distinct differences the skin wasn't purple but we do have purple elements in our image so this checkpoint can handle skin tones quite well providing their natural next I wanted to see what results we would get by testing different ethnicities and there were some variations in the results we got from Asian to African and English with Russian and Native not producing any particularly unique results Indian produce pretty weak results and the faces across the board do look the same so you may need to use IP adapter to help get a more varied set of Faces in your images moving on to testing the age I did notice that the model struggled with age in these tests giving us no age differences between characters I thought that I was going crazy and tested this on a few different seeds with a slight prompt change and we did get an older age in one instance but in another we got a young girl with the others being the same age the results when trying to use age in my prompts was unfortunately inconsistent so you may need to experiment to try and get the results you want then on objects I really love how these turned out we have some really cool designs which are most importantly accurate to what I would expect while having some flare and personality for example look at the sword with the wings in the hilt or the cake with the strawberries and biscuits on top we have accurately represented objects which have some flare and personality driven by our prompts and that's the best combination we could hope for with a model promoting creativity also on animals these turned out amazing I'm using very few prompts but we're getting really nice designs straight from cyber Punk and the animals are recognized and look anatomically correct some like the dolphin penguin and dogs are more stylized while the hamster looks more realistic but across the board it's a fantastic result which can be further worked into what you desire lastly on landscapes these also turned out pretty good across the board I'll forgive the the sorted text in the city piece because most models struggle with Texs and the results are pretty consistent with our other designs the space station is also very good and looks believable to a person he knows nothing about space stations we even have an astronaut popping out of one of the hatches our swamp looks great and mixes with the neon colors well and our Carle reminds me of the Disney logo and looks like it could make an excellent Etsy poster but in conclusion I really like this checkpoint it produces interesting and varied results without being too fussy with the prompts I think that this is the closest model I've found to Mid journey in terms of producing results that are good right off the rip and hopefully with some further development we can have something similar to Mid Journey that allows us to take advantage of staple diffusions open- Source nature but like the video so the algorithm picks it up and let me know your thoughts in the comments and of course subscribe this is bite siiz genius and I hope you enjoyed

Channel: Bitesized Genius

Views: 1,993

Keywords: unstable diffusion, stable diffusion, midjourney, automatic1111, stable diffusion checkpoint, stable diffusion checkpoint models, stable diffusion checkpoint download, Realistic checkpoint, checkpoint realistic, realistic checkpoint stable diffusion, stable diffusion tutorial, stable diffusion video, stable diffusion prompt guide, stable diffusion checkpoint tutorial, stable diffusion checkpoints explained, stylised checkpoint

Id: eAgHlnZvx9c

Length: 11min 19sec (679 seconds)

Published: Mon Apr 15 2024