kegant
workflows (comfyui):
✨ v4 (NAI) ✨ https://civarchive.com/models/1856037?modelVersionId=2100596
✨ v4 (CHROMA) ✨ https://civarchive.com/models/2300447?modelVersionId=2588510
✨ v1-v3: ✨ https://civarchive.com/models/861472?modelVersionId=963859
if you aren't using comfyui, GL!
V4 CHROMA UPDATE: Due to popular request, I have trained the model Chroma upon the v4 dataset. I don't know this model as well as SDXL but I figured I'd release it for public consumption either way. It is drastically different than the previous versions of kegant as this is not an AIO but a raw UNET, which means you'll need the clips associated with flux/chroma. The clip locations of the ones I'm using can be found in the workflow posted above.
V4 NAI UPDATE: V4 is a lot of the V3 dataset, although some images were pruned, but most importantly it is a switch from pony to noobai-vpred as the base model. As such, follow conventions appropriate to how noobai was trained with proper danbooru tags. For help, view some of my published images for tagging styles that I usually put at the start of my prompts. The main reason for comfyui usage is to tone down the saturation and contrast of vpred. While the showcase images are all at full denoise, it is highly recommended to turn the main ksampler down to .8 denoise. The reason for this is due to the way the denoising patterns work on vpred models is much different than eps (or epsilon models). This version is not perfect and likely will need a revision as some tags like blur and depth of field are still a bit of an issue. Use things like 'blur' or 'depth of field' in negatives if it keeps going too strong for your tastes and it should fix it up right for you. If you're stuck using the generator and can't make manual sampler changes such as what I'm doing in the attached comfyui workflow for V4, then putting things like 'red_theme' AND 'blue_theme' in negatives can help, but honestly, if you're using this checkpoint, you probably should just use the attached workflow and see how I am using it.
With that said, none of the showcased images were photoshopped or editted post comfy, nor i2i'd, but I AM using a face detailer. This checkpoint struggles harder than the previous pony ones with far away shots, such as 'full_body' or 'wide_angle' for faces, so a face detailer is highly recommended for this version. There is a face detailer attached to the v4 workflow. Its really not hard to set up and they run faster than a full latent anyway and very much so improve faces from far away. I'm using this guys guide here (and it works great):
https://www.youtube.com/watch?v=gDBeKIa4sHA
V3 UPDATE: V3 is mostly a monsters update, with a few cameos but more importantly it is a finer grain control over artistic elements. I have done some obscene things with this version such as manually editting a lot of the source images by hand in gimp to remove as many jpeg artifacts as humanly possible. Watermarks are non existent and aren't necessary to negatively tag, plants and fauna issue has been rectified, and hopefully males are easier to generate as I added a lot. To see the full tag list of the images I've included in this update its under 'about the version'. Many images in v3 have been tagged with very powerful monikers, namely 'film grain, halftone effect, dark fantasy, muted colors, sepia'. You may see me frequently use them because the source images my prompts are pulling from contained those elements within the source. If you don't wish to see any of those and standard anime, use those in negatives. The style is so strong in some cases that it may bleed if no prompt is issued. Some weapons were added, namely swords, guts's 'massive sword', and katana work (from Cis). Generating images with weapons is always going to be painful no matter how good you are at prompting due to SDXL's limitations, but hopefully some of those added images of katanas and swords will help guide the model into more accurate poses with them.
V2 UPDATE: V2 is the first version I have trained and influenced myself manually. It is still mostly the same stack as V1, however some weights have been lowered and the images I have added and trained fixed a few issues V1 was having. V2 focuses more upon desert style lighting and effects, as well as changed the artistic style slightly for a bit smaller eyes and a bit smaller lips. The lighting has only gotten more ridiculous in this version and I feel if I try to do much more with lighting the whole thing will collapse. I guess we can call this the kegant dune update.
kegant PDXL is a pony based model focused on mutating pony to a more retro and gritty appearance while focusing heavily upon lighting effects.
It primarily focuses around the baked version of 5 separate loras and 1 embedding into the ponyv6 model. These models are:
https://civarchive.com/models/366990/pony-custom-styles?modelVersionId=454703
https://civarchive.com/models/341353/expressiveh-hentai-lora-style?modelVersionId=382152
https://civarchive.com/models/550871/bss-styles-for-pony?modelVersionId=669776
https://civarchive.com/models/122359/detail-tweaker-xl?modelVersionId=135867
https://civarchive.com/models/118418/negativexl?modelVersionId=134583
If you can't see the image the following settings were used during the bake:
Kenva: .8
ExpressiveH: .8
GLSHS: .8
add_detail: .5
negativeXL_D: 1
Concept Art Twilight: .3
Please note that this model has biases in generating females, and prefers to keep subjects not too far away, and not too close. Producing full body may be a little difficult but keep in mind if you specify things like 'shoes', 'boots' or 'feet'/'toes', it will much more be inclined to give you those full body examples you desire. Remember -- This is a pony based checkpoint. It much prefers danbooru style tagging instead of plain english. Sometimes, less is more. Overloading a prompt with too many tags makes it harder for it to understand what to do. If full body is important to you, tag that at the start of your prompt as the higher up in the prompt something is, the more importance it gives it. You can also weight it manually which will help further. I leave all my prompts open on this checkpoint if you are seeking guidance on how to use it.
With that said, this checkpoint isnt as flexible as the goat (aka v6), however what you pay for in flexibility, you gain the additional lighting, artstyle, and speed of generation. To gen the same stack of images with all of the associated loras baked in is about 3 times the speed vs if you did it with v6 and the entire stack, which was the primary focus of this checkpoint.
✨ Please Share Your Cool Creations Below! ✨
Thank you all so much for trying out my first checkpoint.
Please refer to Pony V6 model page for more detailed prompting guidelines.
☄️ Generation Recommendations
* All preview images were generated with no LORA's except for the final two with Haruko Haruhara and Lain as pony has no concept of these characters and they are highly stylized characters that would be very difficult to prompt alone. No other resources were used, just pure text to image with a second pass of latent upscaling only (no pixel upscaling was involved).
Most sample images were generated with ancestral samplers of these types for the initial pass:
Sampler: Euler A / DPM++A
Schedule Type: Karras
Steps: 20 - 30
CFG: 2 - 6
Clip Skip: 2
Denoise: 1
The latent upscaler used was very similar to above, usually opting for a Euler variant as they are typically faster to produce images.
Sampler: Euler A / DPM++A
Schedule Type: Karras
Steps: 15
CFG: 2 - 6
Denoise - 0.5
Upscale By: 1.5-2.0
For some tips on generation, the lower your CFG on the latent upscaler (and steps), the more 'painted' look its going to have, with softer less defined features which creates the 'mist' look in some of the images. Vice versa, the more cfg added, the more 'baked' and shiny it looks. 3.0 cfg is perhaps the closest medium between all the loras that emphasizes each of them the best. For the Harley Quinn image attached, I went nuts with a 10cfg setting to show as well what that looks like, however its highly abstract.
PLEASE look at my attached workflow as it details how you can manipulate kegant to the fullest whether you prefer the sleak glossy design, or the softer extreme retro vibes with film grain effects.
One last note - this checkpoint tends to add 'jpeg artifacts' and various fauna such as 'plants' and 'flowers'. Also it tends to add 'cyberpunk' elements. Add those to the negative prompt if you care to not see those, and it should handle removing them fairly well if so desired. For male work, specifying '1girl' in negative extremely helps it, although the checkpoint as previously mention highly prefers females.
Description
FAQ
Comments (11)
New kegant let’s goooo
important update for v3: please note some tags have been changed, this can be seen in the 'about the version' of the checkpoint. 'side view' was a term i always used incorrectly, when the proper danbooru tag is 'from side'. this has been updated. also check out the hundreds of tags in there for something that might interest you. i had to do some serious sweeping over danbooru for some of the tags because i wasn't even sure what some things were in some of the images. 'obi' for example, or 'kote' or 'yakuta' or 'ascot' are some of those examples that i had no idea what they were (or what they were called) but am keeping up the pony tradition of sticking to danbooru as much as i can when training, even if pony doesnt know about these things.
Will there be an illustrious Version?
I sure don't think there ever will be, they left the CivitAI discord server, saying they wont return and shit. I honestly expected V3 to be illustrious or noobai because hype train and shit.
Exactly what Skyger (Creator of Kegant) said:
" guys/gals im gonna be leaving civitai. i dont wish to be here on this platform anymore. theres numerous reasons to why i don't wish to be here, but i'll try to sum them up:
1) no one in here has posted a single image to kegant v3 that i spent hundreds of hours upon, my own money on runpod instances and manual tagging of images, as well as cleaning up foreign images that dont even belong to me in photoshop the best i could. greathunt is the only exception but no one else in here has. while i realize everyone has busy schedules and it was during the holiday season, its been a month and i was expecting more people who are in this channel to actually ... post something, anything, even if it wasn't that great of an image. since no one has, i don't see a reason to continue this project.
2) civitai is ignoring extreme vote botters on the regular. its actually triggering me beyond belief that they're allowing numerous accounts go go unignored. this further reduces my passion to create new models as people are only concerned with hits and popularity, but its extremely difficult to get peoples numbers up if everyones all either focused upon their own models or they are trying to combat against people spamming buzz bots. rosaline goddess user was the camel that broke my back this evening, a user created 3 weeks ago who had the audicity to post 'finally ... 2k followers' as their update. this user doesnt create models, it doesn't even upscale images. kegant v3 is technically older than this user account. just put that into perspective. chronoknight, a very prominent and hugely successful model creator (go follow them if you haven't even commented on it in modreport, and it was ignored. this isnt even including rodzilla who just posted his first flux gen today and was ... you guessed it .. the number one image on civitai.
3) civitai is promoting heavy usage of illustrious, and some of the models and forks i see of it are extremely not something i want to support. this is the type of crap i cant stand behind, but civitai is: https://civitai.com/models/955971?modelVersionId=1248240 "
Personally I think reason 3 is scummy. Like let SDXL evolve into something good. Also reason 1 is basically hypocritical to reason 2. He honestly said that shit on the first few days of kegant v3 and around christmas/new years. Although the one thing I do understand is what he can't condone on CivitAI which is the link, reasonable for that.
@Arctenox I don't see what's wrong with the link? it's just cute anime girls to me. could you maybe explain something i may be messing there? Also that's a shame.. i really like this model.. oh well
@Monogrim the link had a bunch of NSFW Loli content due to peds, me and a bunch of people reported that content. So it's pretty much cleaned up and by the time you looked another NSFW one got taken down recently. T-T
Nothing wrong with the model in the link, it's the NSFW content that's usually wrong with it, that people posted to it with. Cause some mods said that Lolis even with an adult body depicted in NSFW or skimpy clothing is still a Loli. Like for example model creator put basically what appears to be a fairy child in skimpy clothing, that stuff is reportable and can give you buzz if actioned.
Model itself is borderline if people don't use it for NSFW. Which some people used it for NSFW sadly :(
On a different note I woulda totally used kegant v3 if it had more prompt coherence and was Illu or Noobai, noobai is practically a fine-tune of illustrious and has Danbooru and E621 tags it was trained on. Illustrious is just Danbooru. Another random fact you probably already know is Illustrious is a fine-tune of SDXL.
in a word: no.
i have tried to capture the style that i wanted and even when rigorously training on illustrious, the style was being compromised due to illustrious's bias for lolicon. perhaps it's specific settings within the learning rates that i haven't fully understood but it resulted in failures that i wasn't willing to release.
i believe several people have tried, mimicking it and some people have come close, but the granularity that pony gives is much stronger from what i've seen than illustrious as far as presentation goes.
as for nox's response, i do not make models to follow every prompt imaginable in the human brain. there is no pleasing everyone, but in the future, leaving blanket statements such as 'better prompt adherence' is not helpful in the slightest. rather ask specifically what you want it to do. if its characters, that you seek, theres loras for that. if its specific actions, again, theres loras for that.
finally, if i wanted strict prompt adherence and that was my ultimate goal, i wouldn't bother with pony nor illustrious, but stick with flux as its text encoder is far more advanced than sdxl, so the point you're making, while somewhat valid, is moot given the current models available.
@skyger Darn. oh well. However I do want to thank you for keeping it here and it easily become one of my favorite models to use especially for the retro aesthetic~
@Monogrim thank you i very much appreciate both positive and even negative feedback alike. glad you enjoy it :)
@Monogrim 4.0 is a noobai-vpred version.
@skyger LET'S GO~ Thank you!