<h3 id="version-2.0-.-semirealistic-model-that-can-pull-from-cartoon-source-and-still-maintain-fairly-high-realism-and-with-good-details-(slightly-cooked)-and-well-at-least-decent-lightning.-2c54tq73l">Version 2.0 . Semirealistic model that can pull from cartoon source and still maintain fairly high realism and with good details (slightly cooked) and well at least decent lightning.</h3>START WITH NEGS PROVIDED IN SHOWCASEIf you want very realistic this model easily takes on more realism input, but I leave that choice for you. That said I think this model is at its best with 1-3 more lora on it to really fall into a nicely defined space. I almost always use different loras ontop of this model so this is like a baseline.If you don't know what to start with I can recommend 80sFantasyMovieV3Pony, its sweet 10-60% smth like that. Or Disney3d lora for realistic disney.CFG 3-6?. I advice to run with 3d and realistic in negative but its not needed. It can pull from realistic and does it pretty well, but I think it does its best from cartoon source which is guaranteed if you put realistic in neg. remade / different base but core idea is the same. salvaged a lot of the "freedom" and variance that had been lost. As of now it pulls hard realistic but it also takes a lot of input from cartoon source. If you want cartoons now you can add DST-Realism at -30% and go from there probably need a few cartoon loras to totally let go of realistic input. I might upload a more toony version later if it can surpass v0.9. Note that the variance comes with ugly pictures at times and at other times you get things that is good that you would never get with a model that is more strict. Zelda in this model can have many different faces and I like that a lot, it gets boring with same face/similar pose on every gen in my opinion. This model should put out high variance in poses.Check sample pics for negative prompt to start with, then you can remove as you feel necessary. 2.0 also needs a bit of tags to get going with color and composition, you can't really prompt only 1girl and get good results on this one. One of the models I merged with is cursed with wolves :D so if nose/area gets funky try adding wolf to negative, this is especially true for furries although I put so much realism in this one Im not sure it does furries well. I did make a great judy hopps though with slightly different parameters but it needs lora help for sure either way.As of version 0.9 all suggested tags below are just recommendations. I leave it, as they are still needed for earlier versions. The model v0.9 is well balanced and can do pretty much anything with prompting that you would expect. If 3d and realistic takes over just add any cartoon lora and use Euler A.BadRomance is a model that tries and fuse romantic and dystopic style while trying to bring a somewhat original face and body to the mix. I think I did fail with making face truly unique in the end, however I rather have a good base that is easily inpainted upon so one can choose their own face rather then being forced into something that I fused in.My only recommendation is to use "glowing eyes" in neg if you gen on low resolution or far away shots / or inpaint with this neg. And if you are on low CFG try "pastel" in pos. "Long nails" is also very helpful to make better hands on low resolution. START WITH CFG at 2-3! (4 is very stable but its better to start too low then too high)Showcase are as is, no inpaint, no adetailer, no hirez fix with 4 exceptions, they are marked with inpaint or img2img and the main showcase zelda img is hirez with img2img but it was fixed with the model itself with Euler A. If you like style of main showcase image its better to check in <s>CyberRomance</s> now called DragonSlayersTavern model as that has most of that lora stack in it. However this model is much better balanced and more versatile.Scores are not always needed when using realistic tag. Note that scores can reduce realism. Sometimes it is very much needed though, experiment.To be honest after printing a cool vibey sketch of dungeons and dragons gen with only "dungeons and dragons" in the prompt I don't think I will touch this model anymore. To me that was highly unexpected and very cool. To me that speaks of good balance. I don't wanna ruin that. The model will continue under <s>CyberRomance </s>DragonSlayersTavern.<hr />Highly recommended Positive: RealisticOptional: Pastel, color, detailed eyes, realism, photorealistic, animeHighly recommended Negative: 3d, glowing eyes //// For anime put realistic in negativeOptional negative: Wrinkles, ugly face, realistic, monochrome, greyscale, glowing eyesRealistic: DPM++2M Karras / Anime: Euler A or dpm++2MK (Euler helps pushing away from realism, I prefer karras)CFG 2-7 (4) (if no color its better to change prompt then CFG, try keep CFG low)steps 20-30Bad Romance is fitting name. It will be bad at times and at other times... well try it out and see. Showcase pictures on v0.8 are as is, no hi rez fix, no facedetailer. It has access to all sources included in base pony. It doesn't hold your hand though with a strict style, you gotta know how to prompt. And when to add some loras to help on the way.Resolution - The showcase images are mostly at higher resolutions. Model works best at higher resolutions but can also make decent pics on lower but hands and eyes can be difficult to manage without inpaint on lower. Showcase pics are not enhanced at all, its the base gens (unless clearly specified otherwise)."Glowing eyes" in negative seems to help a lot with making better eyes.If you want a prompt just ask, I don't hide it, but I use Krita and it doesn't transfer metadata automatically, so its bothersome extra work for every pic. Just ask and I will add it for that pic.Refining on realistic models seems to work really well. So why not just use the realistic model to start with? This model is MUCH more flexible in its output and will give MUCH more diverse gens from a prompt I'll dare say. Have you seen the stale backgrounds that are on most realistic models? My backgrounds are great. Why? Cause nothing has been hard cut out from base ponyv6 model. As of version08 you can make images with open prompt, no style tags and no neg tags on style. Model has wide source of material. For example girl falling into pool. Then you can refine with tags you want. THEN you can refine them on another model if you want. A lot of other models has cut away so much that this more "weird" material does not exist in them. You need therefor to leave style open when attempting this cause if the style doesn't have it you simply wont get it or get bad gens. For example you might just get a sketch of a girl falling into a pool. But that is good enough material to keep working with.Model is flexible enough to accept lora styles. So if you have trouble with staying in anime style, add an anime lora at like 30-60% etc.<hr />This is an attempt at making a model that is fairly close to base pony (I like models that are, cause they have more variety and are less rigid) that changes the base face shape somewhat while also trying to maintain a more natural body type<hr />Below is mostly in regards to version 0.5 but also about my vision about this model as a whole. Only autists and nerds (like me) might find benefit of reading it.IMPORTANT NOTE - All lora values discussed below are in regards to both strength_model and strength_clip. For example masculinity slider is mixed in at strength_model -0.5 and strength_clip -0.5 so to get back to base you would add that lora at strength_model+0.5 and strength_clip+0.5. The same applies to my prompts if you are on comfyUI you gotta make sure your lora input changes same values as mentioned above (both clip and model strength for the loras used in prompts) to get same results as I do. (whether or not comfyUI does this if you prompt the lora I dont know, if not you gotta use a lora loader node). Links to these loras at bottom VThis is an attempt at making a model that is fairly close to base pony (I like models that are, cause they have more variety and are less rigid) that changes the base face shape somewhat while also trying to maintain a more natural body type. I would say a more realistic face and a more realistic body however make no mistake the style is not realistic. That means the girls are not perfect 10/10 with perfect lips, hips, boobs and pointy nose in every gen and that is the whole point. I would describe it as some kind of oil paint. To avoid getting the classic pony face and body this was balanced without using scores as scores heavily steer body and face toward a certain standard. You can use scores in this model if you want to ofc, its needed on some prompts for sure. On others its best to avoid it. Experiment with that. For the original style that was intended it will go away with scores but some gens look terrible without it. If you use realistic tag using scores can make eyes even worse keep that in mind. The lora mix in this depends on pulling from realistic source to come out as intended and not from 3d source so that should be in neg, but probably there are cases where not having 3d in neg looks good to. The idea of the model is preserved even if you want less realistic ouput, simply put realistic in both positive and negative, you can also add cartoon, anime etc in positive, together with realistic can probably be nice and with cartoony loras. But I didnt make the model to be used without realistic tag. If you do not want to use realistic tag in positive you should use score tags instead. the normal ones. score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up. Who knows maybe it can make some nice stuff with that?The style I personal seek from this mix is in the CFG zone between colorful and a total mess. CFG 4-5. can be lower can be higher, depends on prompt. I had prompts that can easily go down to 2-2.5 area for that really dirty muted color look that I like. When it feels "cursed" its close to what I want. Ideally you wanna be in the zone that gives cursed gens and gens that look alright. The style changes completely in that zone. A gritty dark theme was intended for this model but it also does romantic cute stuff. And if you want more cute just turn CFG up a bit and use for example pastel tag.I have two pretty degenerative loras in this mix which I for a long time thought about if I was gonna include or not in the model or just recommend to use. I ended up including them. Its masculinity slider at -50% and lowlight lora at +60%. Masculinity slider is extremely degenerative so if you get severe problems with limbs, hands and mutated gens you can try adding that lora at +50% to cancel it out but another way is to just increase CFG. Samething for making cute, add 50% masculinity and -60% lowlight to go back to a normal light / detail base for making more clean cute images. Some of them are included like this in examples, the prompts highlight this.In my original mix I had even more degenerative loras in the mix that I left out. Ass adjust, penis size slider and body type slider. All of these have degeneration on the style on a mix that is already as low mixed and fragile as this one. So if you want even more of what the model does then add these and experiment with the settings. Turn up CFG or down as needed to get back in the golden zone of degeneracy. And if you wanna go ham with penis size for example then you might have to turn up masculinity slider to its base value (adding +50%).This model does fat, chubby, saggy etc really well in my opinion in a dark gritty setting. There is enough detail left for plenty of wet and cum effects etc. So you can make some pretty nasty stuff with this. But hey do what you want. I left the model pretty "open" its easy to add loras unto it and change the vibe a lot. Keep in mind the loras included are not mixed hard into it so most likely you can use your usual loras at lower strength. And some loras will not work well here cause they can only to their job at 100% and they will ruin the style of this mix at that value. Apart from already mentioned loras expressive is at 50%. 40ths romance at 50% (this one changes the face and is also responsible for eyes not being great, sadly). But I'm so tired of having the same faces in pony it was a sacrifice I was willing to make. There are a few other loras in there to act as contrast enhancers to weigh up against the degenerative loras at low values. The fact that everything is mixed in low gives plenty of room for you to do what you want with this model, you can change style a lot. The downside is that a charachter lora can change the style a lot. Where other models are so hard loaded with loras that a charachter lora on 100% doesnt really dent the balance. It will dent the balance here a lot. But hopefully after reading this you will know what other loras to add to get the degeneracy back.Also yes its not perfect, but as far as I know no other model provides what this provides. So if you want refinement just img2img in another model that is more stable. Or inpaint a bit. Might sound like a bad excuse but there really isnt any other way for me to achieve better what I wanted with this. Maybe in future versions I can fine tune a little bit to get better eyes and stable anatomy but every way I know off kills the style I want here so it is what is. If anything I will make future versions even more "loose", reducing some loras and add others to compensate so ponys vast material comes through even more. One of my main goals here, to make a prompt yield many different outcomes and not always the same. Not rigid. Eyes and hands is my biggest annoyance here and after that I wish it was even less rigid than it already is. Its gives pretty mixed gens but I like it to be even wider spectrum for a prompt then it currently is. I also do not like on relying on 50% expressive. It gives problems with nose and also hair and skin color. (makes hair brown and skin whiter and nose more pointy). In future versions I hope to find a lora to replace this with smth else or at least being able to reduce it.Actually upon further testing anatomy is good here, it just needs high resolution to print good hands, nose and eyes.<a target="_blank" rel="ugc" href="https://civarchive.com/models/486786/pony-xl-leco-masculinity-slider">Masculinity slider https://civarchive.com/models/486786/pony-xl-leco-masculinity-slider</a><a target="_blank" rel="ugc" href="https://civarchive.com/models/432511?modelVersionId=481826">Lowlight https://civarchive.com/models/432511?modelVersionId=481826</a><a target="_blank" rel="ugc" href="https://civarchive.com/models/465379/penis-size-slider">Penis size https://civarchive.com/models/465379/penis-size-slider</a>Ass adjust: <a target="_blank" rel="ugc" href="https://civarchive.com/models/477074/ass-size-slider">https://civarchive.com/models/477074/ass-size-slider</a>Body type: <a target="_blank" rel="ugc" href="https://civarchive.com/models/468146/body-type-slider">https://civarchive.com/models/468146/body-type-slider</a>romance lora: <a target="_blank" rel="ugc" href="https://civarchive.com/models/380531/1940s-romance-style-ponyxl">https://civarchive.com/models/380531/1940s-romance-style-ponyxl</a>There are a couple of more loras added at like 10-20% each, I'm gonna hold on that recipe for now.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1ac262bb-b930-477a-8b0b-8ed0f6c3ea21/width=525/1ac262bb-b930-477a-8b0b-8ed0f6c3ea21.jpeg" />

BadRomance - Pony

Version fixed hands and eyes. Good output without facedetailer or hirez fixing. Better background. I added 15 loras that has realism as base at 5% (some of them 10) each. Thought it would affect face more then it did but probably good it didn't. This pushed out anime and cartoon style from the model and I dont want that. I added 2 anime loras (smooth style 2 anime at 10% and Vivid neon mess at 30%) to compensate for this. The effect on the semirealistic style was positive which I did not expect, helped fix my background problem. Reduced lowlight lora by 30%. Reduced expressive by 20%. It did however slightly push face closer standard which I want to get a way from a bit more, maybe in a future version who knows..My vision right now is not to ever make this model super realistic.To fix: Head is slightly small and shoulders are a bit broad.

Files with SHA256: 894a251971f1e4242f407b2dc9f59e4eadd9d528b9de7f0537a3e88c6d25f128