Before you enter please read the rules so you don't get lost, if your new to the model its a good idea to at least read the 1.7 sections (green section VV) colored parts to have a good time and good results and I would start by using 1.7 its the one I know I like the best so far:

1.9 - I think this is an improvement but the reduced realism and potentially worse base style might not be.

It is more flexible though, might need one lora added "to get there". Teeth can be uncanny but I think its been like that for the latest versions. Hands, backgrounds, adherence, I think most things improved on 1.9 but its less "blingy" on first look maybe? Anthros and furries tend to become a bit wolfie and dog like since around 1.7 or smth. Adding wolf in negative can help with this. Its kinda the point though to not lock the model to much in one style but let it be flexible, so potentially while not as cool on the surface as 1.8 this one is a stronger model I think. Need to test lots more and then I will decide how go about v2.0

If you want higher realism go with 1.8 if you want maximum fantasy, adherence and overall a great model go with 1.7. And if you want even more quirk and color go with 1.0

This model is weighted differently from other models, copy pasting prompts can work and give great results but depending on your prompting style it can also give very bad results especially if you use your standard sampler settings. You gotta approach it with somewhat of a clean slate and build up your prompts with some patience before seeing good results. And for some pics choosing 2.5CFG will make something great that looked meh on 3-4 CFG.

All images in showcase for 1.7-1.8 are base gens, no upscale, no face fix, no refinement even within the model. So they could be better, but yes I choose the good ones. Except the fire wielding goblin/orc picture. v1.8 quality increases MASSIVELY run with latent upscale see pics posted and pinned in posts.

Welcome to DragonSlayers Tavern. DragonSlayer is said to have the best tavern maid elves in all the kingdom, some even say the whole world. This is a (semi)realistic pony model with some added bling bling that is still totally ponyv6. Its goal is to bring better fantasy, clothes, races, backgrounds and darkness, then most other realistic models. And to provide furry kinds that are not uncanny yet fairly realistic. Here we welcome all races, elves, humans, orcs even those furry things. One of little imagination might wonder why go through all this trouble to have good monsters and elves in the same model hmm?

V1.8 - not majorly different from 1.7, but has a clear advantage in realism in some kinds of scenes. slightly more realistic. more adherent to ethnicity. For realism a light lora that brings red colors is advised, like sinfully stylish v0.1 at ~30%. I didn't include it in the model cause it will fack up other styles but without this realism easily looks bleak and stale.

I actually killed off more cartoon base then I first thought with 1.8. As such it cannot do some action shot things that 1.7 can. If you want more pinup girl realism 1.8 is better. If you want more dynamics and make like laser sword fights or blood splatter and cutting a dragon in half, with red sonja lora or similar I recommend 1.7. If you want DnD in a more cartoon style try out 1.0, 0.2 or Badromance 0.9

"Less magical" but can get that back in large part with reel mech lora at like 50-80%. Also almost all loras work with this model, cartoon, anime and realistic based loras. So you can change colors, light and style A LOT if wanted. 1.8 is less interesting in its base style but it can be pushed either way. 1.8 has bigger need of loras to help push style where you want it.

To get desired style using proper style tags in both pos and negative is needed for best results. It can make cartoon style but needs persuasion by tagging. Other then that much like 1.7. Slightly worse hands sadly. Eyes are worse on distance but better close. So 1.8 is best for higher resolution workflows. If you can't upscale 1.7 is probably a better idea if you want anything beyond a cowboy shot.

1.8 no longer needs refine at all on dpm++2 karass unless you gen at 1.1-1.5CFG, but will improve with it ofc. For Euler A - hyperCFG lora at ~40% can do lots of nice things and its almost needed for Euler A I would say now, because model is less cooked Euler A looks more plastic in base so I like to add this when using Euler A. Links in 1.7 section

1.8 handles higher CFG better I think. Its harder to get great monsters on low CFG though but still possible with added cartoon/pony_source tags as this is the source with good monsters in it and realistic and 3d in neg.

If you want hyper realism I recommend refining on YAPR at like 5-40% depending on how much of that style you want.

I made my first lora to stabilize my vision of realism in this model without burning it into the model and ruin every other source. Its found in resources DTS-Realistic-Realism strength 0.1-0.7. It helps shave off 3d source when it shows its face to much. Currently only tested on females and it seems its fantastic on some scenes and terrible at others :) Worth to play a little with I think but it probably has lots of problems. I think I just have to use a custom lora stack instead when I want this.

///1.7 version - massive improvement

Try reading at least the colored parts of this section to have the best time with this model

This model brings a slightly cooked karassdpm++2M result with good clothes and backgrounds. The added cook adds detail. Which makes result after refine in realistic model of choice very good. So why? Because its much easier to refine one time in a realistic model than to use realistic model from scratch and having to inpaint better detail, clothes and backgrounds. Also obviously this one makes great monsters, furries and all that. Its easier to refine extra cook and details away then it is to add detail. It is especially noticeable on glowing things but it applies to everything and on low CFG it makes very interesting results as can be seen on my "fire wielding orcs".

I just tested adding more hyper lora and it works really well. I don't wanna mix that into the model as its only useful on Euler A. But I do recommend you test it like in the main showcase picture together with sinfully stylish. Makes for a very nice blend.

https://civitai.com/models/800496 hyper lora

https://civitai.com/models/340248?modelVersionId=380971 Sinfully stylish I like 0.1 the most.

!!!! CFG !!!! start with 2-3. 1.1->4 is the models best area. ye crazy

This version is more realistic, better at girls and humanoids in realistic style. It also lost some of its high fantasy elements and a little magic, but the model has anime and cartoon base in it so it should handle almost any lora. Experiment :)

Start with Euler A at 2-3CFG. Latest version seems to handle a bit higher CFG. Note that Euler A easily gets plastic and only reaches semi realism. When you are comfortable with model on Euler move on to dpm2++ karass and check it out around 1.5-2CFG its another world down there it can take realism much further. And after that creative SDE karass. Karass can get a little cooked at times, refinement on Euler or some kind of realism lora can shave this off. I don't mind leaving it like this. Cause it gives user options. Sinfully stylish v0.1 for pony does the trick well at 10-30% to round out the image when not genning on Euler A.

Positive prompt quality tags: Dungeons and dragons, score_9....., vibrant colors

Only negative needed are: Glowing eyes, Long nails, Inverted nipples. and the common age tags ofc. ugly face and/or wrinkles if realistic tags bring out milfs that you do not want. If you want long nails be prepared to gen on HIGH resolution / upscale, inpaint. Other good ones: Goo, 3d, anime, cartoon, realistic,

A long negative prompt that you usually use might mess up the model. It is sensitive use negative prompt to save good gens and fiddle around until hands get good if needed.

Always describe colors first. "Do not touch CFG" If you get greyscale/monochrome it is better to describe the colors of the scene and add/or "vibrant colors, pastel", and try adding scores, etc then to place monochrome and greyscale in negative or up CFG. As that source is needed for well rounded images. And CFG impacts style quite heavily especially on karass. But sure some pictures having these in negative will bring out more color if that is needed. In general it looks most realistic with these not in negative.

Scores are not needed, sometimes they bring a lot sometimes they ruin realism. Experiment!

In difference to other versions it doesn't need refine at all. Shit looks good on karass and Euler. Well on Karass most times, it still looks best refined but can also be helped with loras. For monsters you need to go deep down on CFG though 1.1-2 (3maybe) area if you want them scary, otherwise they look like world of warcraft. Hands are not perfect but they aren't a total disaster, a few gens or an inpaint and they are set. Of course refining still makes it way better so

Model now has a comfyUI workflow attached displaying how to refine it with itself. (plz excuse my noob workflow I barely use comfy). Its not really needed anymore but it will improve pictures and maybe its needed for the most realistic images that you would gen on creative sampler. If you dont have comfy the process of refining is explained further down. For higher realism refine with a hyper realistic model. Why not use a realistic model to start with? My model has better clothes and fantasy elements both medieval and futuristic and it has better adherence with complex stuff, like jumping and holding things (and that goes for porn as well). Some backgrounds are also really good here.

Note that while this pulls toward realistic and semi-realistic, it still has cartoon and anime source in it. So you can do lots of things with this model. Be mindful of your tags to get the style that you want. Base style of Euler A is semi realistic / 3d, if you want realism. (3d in negative - optional) and lower cfg and not using Euler A can all be helpful, with realistic in positive (optional but it helps and keeps style stable over prompts). Creative SDE sampler is the most realistic but needs refining. Cartoon and anime can be a good idea in negative.

While 1.0 still has some kind of special "magical" vibe to me a lot of that can be brought back into 1.7 by adding reel mech lora at ~0.5.

Welcome! Enjoy your stay.

VVV Recommended loras below VVV

You may do whatever you want with this model except sell it, that was a prerequisite of one or more of the models I merged with so no no on that. Otherwise feel obliged.

///1.5 version----

The model now has a comfyUI workflow attached displaying how to refine it with itself. (plz excuse my noob workflow I barely use comfy)

All images in showcase are base gens, no upscale, no face fix, no refinement even within the model. So they could be better, but yes I choose the good ones.

Only negative needed are: Glowing eyes, Long nails, Inverted nipples. and the common age tags ofc. To not get a total milf here: "ugly face" and/or "wrinkles".

!!!! CFG !!!! start with 2-3. 1.1->4 is the models best area. ye crazy

Leave the rest of you negative luggage at the door cause it will mess up the model. You can of course add negatives later but start clean and with suggested negatives or you have no clue. Hands are bad enough and does not need to get fucked over even more by your bad hands, mutation neg prompt.

Always describe colors first. Do not touch CFG! If you get greyscale/monochrome it is better to describe the colors of the scene and add "vibrant colors, pastel" etc then to place these in negative. As that source is needed for well rounded images. But sure some pictures having these in negative will bring out more color if that is needed, but you get kind of less shadows.

Euler A for good ez gens. But it likes to pull from 3d source on more complex prompt. And its not very good for monsters, dragons etc cause it pulls from 3d source easily which IMO is ugly for monsters and dragons most of the time. Euler A doesnt need refine (even if it of course gets better from it) but its more plastic.

For monsters, grit and more realism DPM++2 Karass is recommended.

For even more grit and realism Creative DPM++ 2 SDE etc.

Karass still needs a little refine.

Creative needs a refine but it pulls from the most realistic source and gives the most details making it best if you are gonna refine in another model.

This model is for lovers of gritty, cause it was love of grit that made it. Dirt, grit, dist all that fuzzy stuff. The model is especially good at making dungeons and dragons style pictures in semi realistic / (cartoonish style) - as of 1.5 cartoon base is harder to utilize but its there to some extent still. Along with other kind of fantasy like warhammer, cyberpunk etc. Essentially a base genner that can then be refined into more realistic style if wanted but the style I think stands well on its own. To me what it gens is just magic (at least 1.0 I've yet to decide about 1.5) and the results refined are in my humble opinion much more interesting then what the realistic models puts out. And of course it makes dirty porn as well as it does dragons and warhammer, its as good as it is as -"base pony" is with a good lora stack- at porn, great adherence etc. Ez gangbangs - which more realistic models can struggle with. However problem in 1.5 is it easily reverts to 3d source for these kind of more complex prompts. It can be managed with sampler choice and prompt, its better to sample from cartoon and realistic then from 3d, keep that in mind, For DnD pics put 3d and realistic in neg and cartoon in positive.

Purpose? To yield great diverse base material with good prompt adherence. It can stand on its own or it can be refined. Its not an all easy model to handle (badromance model is very easy and forgiving). Sure you can easily get pretty nice results with Euler A, great results even. But you gotta know what CFG and samples does and when to change them. You gotta know your tagging.

END OF RULES for 1.5 - WELCOME!

Ver 1.0 !!ATTENTION - INSTRUCTIONS!!

Examples of results of realistic workflow is pinned.

Good tags: Dungeons and dragons, realistic, steampunk, cyberpunk, gears, joints, cables, neon, orange theme, vibrant colors, pastel, color, medieval, victorian, baroque, futuristic, scary, black background

If image is grey, describe more colors, add scores, and it will come. DO NOT INCREASE CFG to get color unless you tried everything else first.

Recommended negative tags: 3d, glowing eyes, inverted nipples, long nails

Optional negative tags: fog, foggy, realistic, | anime -> (if you get the babe doll face) - I gotta sort out background for version 2, less fog and more realistic.

Scores?? are not always needed. Scores can fuck up hands and eyes. Sometimes they are very much needed. Experiment!

Some Flux prompts work very well for some reason.

Option 1 EulerA *** - quick workflow

(works well but will not yield the best monsters, armors and is not as realistic on skin etc) Some images in showcase are made like this. Some of them have an extra facefix inpaint added.

EulerA cfg 2-3 steps 20-30, some prompts need 5-6CFG.

It is still a good idea to read the next sections to understand how to best replace face with race of choice if you can inpaint / increase its detail. And also how to get the most realistic images. You can generate on Euler A with realistic in neg to get the more diverse cartoon based material then you can Img2img and move realistic tag into positive to get more realistic result.

For even more realism refine with smth like DTFmix - hyper (SatPony works great!!!)

As of now the model seems to lack a little red tones on more extreme light scenes. RealLowlight lora is one way to solve this (at night, in lowlight tavern, on beach in the sun, etc). Another is to refine in another model.

Option 2 DPM2++ karass - ***** workflow for gritty realistic images

This is how I made the pinned warhammer picture.

Images are slightly cooked that is intentional. In the tavern we bake and cook all day so the images that are generated are slightly cooked n baked. All images in showcase are refined and !!genned at 1.5x resolution!!. Here is what you do with them (V quick version in blue V)

1. Generation - Sampler DPM++2M Karass CFG 3-4 (2-6)

2. Inpaint part 1: When generation is complete fix any potential defect with the same sampler, hands, ugly clothing, facial blemishes etc. At this stage you can add stuff or remove with inpaint using same sampler. Change seeds if/is needed.

3. Inpaint part 2: Delete prompt. Enter race, facial features and facial expressions. (nothing else in prompt or the race wont apply - also realism is increased as the AI no longer is constrained by background) !Change seed! -> Inpaint the face with EulerA CFG 2-6, steps 20-30. 10-50% denoise depending on original CFG and how realistic you want it. Lower CFG here increases perception of blemishes on skin which might offer a better realistic feel. Maybe this is best done with original sampler if you are gonna upscale it.

4. Refinement - Load up the original prompt and

either:

a) invert selection and inpaint/img2img the rest of the picture with Euler A CFG ~2-5? steps 20-30? Denoise 10-40%

b) select areas and inpaint them 1 by 1 this increases detail which might be wanted on jewelery, clothing, nipples etc. And if you have a shitty GPU like me this is the best way to refine after upscaling. I think it actually goes faster refining sections of the picture then doing it all at once. My gen might just fail after upscaling if I do the whole picture. It is also safer, how much does it not suck refining a picture on a shitty gpu for minutes, only for the stomach to have 2 navels? Refining in sections also reduce the risk of this happening at all. And you can refine a few times for each section (switch seed) and choose the best result.

4.5 IF you want to upscale do that now and you can do step 2-4 one more time if needed. step 4 is highly recommended to do after upscale.

5. Higher realism (optional)

Pick a (hyper) realistic model preferably with good background material. (you can skip step 4 if you are gonna do this I think. But I think the double refine might work better still but sure its extra work, make sure to load the original prompt though.) DTFmixHyperreliasm. Refine at 10-65% (25-35 good area). Check model info, I think creative dpm sde3, in general is best but it varies. there might be better settings for this. yaprmYetanotherpony_v90reboot is also good. And I tested refining in an SDXL model today (araminta experiment cv6) it worked well. The seed nr was represented very similar which helped.

Recommended refiners:

SatPony (I only tested cinematic so far and it was GREAT, don't even need to upscale or run at super high resolution) Euler A seems to work best here.

DTFmix - Cute and fast. Does the job well, handles backgrounds which Im not sure satpony does that well yet, will test more.

yaprmYetanotherpony_v90reboot - Fantastic but tricksy. Its made for fatter women and face pulls HARD in a certain direction, but it doesnt really matter if you know ure inpainting well. I was able to actually write letters with this one which was very cool but I think all pony models can do that to some extent. I used this one for the warhammer pic that is pinned. Satpony does amazing on the base picture as well though.

you can of course refine more the once. SatPony first at ~30% then yaprm at ~10%. yaprm has the highest addition of realism but impacts face a lot so dont wanna go to hard with that especially on the face area.

Option 3 **** - using external model for refining and quickly get results

1. Generate with dpm2++Karass (euler A also works) as step 1 above.

2. (optional - (correct and fix face as explained above but with same sampler if u wanna bother with inpaint, if u have facefixer it should probably go after refinement.))

3. Img2img refine in realistic model of choice with Euler or DPM++ sde, at ~2-5CFG and ~20-30 steps. (change the seed)

4. Note that image will probably look overly cooked (and shabby) and with too much details until refined! Especially if you are using Karass as the genner.

Recommended refiner

DTFmix

SatPony (doesn't need face fixing u can refine straight away) and facefix after if wanted. Refine with dpm++ SDE creative, or Euler A seems to work best.

As an example look at the Kawaii girl in one of the pinned posts. I started that with Euler and anime tags in pos, realistic and 3d in neg. I then refined with realistic tag. Upscaled and then refined in a realistic model, then I refined/inpainted only the face one more time. (DTFmix).

----

Option 4. Quick Option within the model ***

(just genning with Euler A is obviously much faster but it easily makes more 3d/cartoony pictures and doesnt offer same grit in end result) Most images in showcase are made like this with an extra inpaint on face (increases accuracy and detail) after the refinement.

Gen with DPM++2MKarass 2-4CFG - 20-30steps, (euler A also works as base genner it gets more "clean") -> Refine whole picture (img2img) with Euler A cfg 2-6, steps 20-30. !CHANGE THE SEED! at around 20-30% denoise This way is faster then option 2-3 and it is enough on some images but if you want the best quality follow steps on option 2 above. Likewise if you want another race for example asian, it can easily get "lost" in complex prompts and you would need to follow step for inpainting the face.

---

END OF RULES - WELCOME!

Why not just use dtfmix/SatPony? Well you go and try make a scary dragon on that model and see. Or try my warhammer prompt on satpony. This model offers grit, dystopia, cyberpunk, romantic detailed themes and for some reason bad ass gens. The cartoon base is still here meaning it pulls from that to make ure dragon and backgrounds (and not 3d source Urk... at least if you have 3d in neg which I almost always do). Same thing goes for everything else. So this offers in my opinion MUCH better base material to work with. Of course if all you want is nude babes go use DTFmix its a great model! (or use this one with black background in pos). But if you want gritty dungeons and dragon themed pictures. Like a dirty dungeon, or dark, moody lively tavern, or a gatlin gun wielding warhammer girl. then I think my model is a better start.

If you don't like my body mod (and I don't always myself) then download 40s romance lora and pull it down by 40%. -0.4 on both clip and model strength. This lora adds more then that though so only if you get a saggy lookin ass or smth you don't like.

Do note that some scenes heavily pull away from realistic and its because of the background or smth along those lines. Which means "black background" is the best positive prompt for getting the highest grade of realism. Adding realistic loras, of ppl or similar or places is enough to push it into true realistic by my standards and then the background falls into line it seems, I have to investigate more. It doesn't matter that much though as refining will push it into realism anyway.

I guess what is lacking most is realistic background material. I think that is a heavy strain to force into the model. Would take lots of room from other things. Im not gonna do that. But if I find a few good loras for this I will recommend using them. For the most realistic prompt "black background" or perhaps "empty background". With some clever img2img Im sure this can work even for more complex compositions. Do note that some scenes heavily pull away from realistic and its because of the background. So be wise with how you compose the scense with ure inpaint tools to get the most out of this.

Yes the model is cooked at around 5-7 CFG+, who cares?

Why did I overcook it like this? Because the added hyper and turbo loras increases detail and noise and just brings out some good looking stuff without sacrificing the base material pony has. This adds more stuff in picture for AI to work with, it makes results on low CFG better and its only down there monsters look really badass. It makes creatures and armor cooler, scarier, weirder. It also makes skin and hair more realistic.

Model has pushed so far toward realism it is now only good for semi realistic and realism. Making anime pikachu on this model isn't easy. It is possible though with cartoon loras. And it can do some sort of semi realistic anime. But if you want cartoon style then there are better models (like badromance which has vast toon based material - badromance even has sketch-monochrome left from pony and you only know what that means if you are a true degenerate). Cartoon/anime based loras can also add to the vibe here. Cr33p by vixon for example works really well for an even more dystopic vibe. And its no problem to go into cartoon style and put realistic in neg here if you are gonna refine in a realistic model later (if you want realism that is). You can also of course refine within the model. Remove the added cartoon loras, put realistic in positive and refine with EulerA. But it makes little sense to use this model for toon style, other models will do that better. With 6gb limit if you put smth in smth else must go... Which is why badromance I think is a great model, I only cut the true shit out of that one I think... It makes cool gens on extremely simplistic prompts in sketch style and above, and can push those gens to semirealism on its own. Only true enjoyers of gritty stuff will understand I guess...

VERSION 0.2

LOW CFG!!! 2-4 area. There is lots of cool funky dirty stuff on lower CFGs. Having to high CFG spoils the vibe, especially for realistic women. More ok for a vespeon. Don't be afraid of exploring between 1-2 CFG area if you want gritty dirty cyberpunk.

DPM2++ Karass

Scores are not needed, sometimes they make things better other times worse. Experiment!

Problem: likes to make greyscale at this time. DO NOT increase CFG to combat this, first try add color words, describe colors in scene. Then add scores. Then you can up CFG if nothing else works. High cfg = COOKED. Sometimes foggy...

Positive: Color, pastel, vibrant colors, realistic, photorealistic, realism, ((anime)) ((retro anime))

Negative: Monochrome, greyscale, 3d, foggy, fog, glowing eyes

Showcase images are as is most genned at 1216*832, no hirez fix, no inpaint, no adetailer. Which means you can pull much more quality out of this then what is showcased.

Also this isn't a merge, its just a 35lora stack on ponyv6, same as BadRomance model.

Semi realistic cyberpunk / armor model. I think it can become pretty realistic if you add a lora based on realism. (80s babes for example). But otherwise I'd say its semi realistic.

Based of BadRomance but this model takes no consideration to other sources than those that are of realistic nature (semi). I made this into another model because my vision for BadRomance is quite different. It might be able to do other sources, I haven't even tried. Versatility was sacrificed for a bit more realism and more mechanical parts.

At first I thought just slapping a lora stack on BadRomance would suffice but some of the loras from early models caused problems, and the lora list for making what I wanted become so long in the end it was better to make this into a new model series. I removed some of the input from these.

For future versions of this model series I will make the model in one go from baseponyv6 as that should increase quality a bit while maybe also adjusting somethings. Maybe I try push realism a bit more, maybe not. As it is adding realistic loras seems to do the job and I think its better to leave it open to the user what realistic lora to add then me forcing it in making it more flexible for user.