V3 is here! Improvements include better eyes, feet and proportions. She is close to V1 in terms of overall body and face shape, while still retaining some of the flexibility introduced in V2. Colors are pulled back somewhat. Use my showcase images as a springboard for your ideas. Some features may be stubborn, such as her tail rings and in some cases: the pendant and choker. Generate a few more images and these will most likely be gone. Same tags as V2! Tags such as "nude" and "completely nude" will help to get rid of her accessories if you're having trouble.
Feedback and requests is much, much appreciated!
V2/V3 Trigger word: ChickenKrystal
V2/V3 tags: furry female, two-tone fur, white fur, blue fur, blue hair, aqua eyes, fox ears, fox tail, animal nose, sharp fingernails, clawed feet, small breasts, strapless bra, beige loincloth, tribal tattoo, hairband, circlet, choker, pendant, shoulder armor, bracer, tail ring, anklet, sandals,
Other tags used during training: from side, from behind, from above, upper body, lower body, full body, portrait, cleavage, staff, holding staff, the pose, ass,
Useful tags for accuracy: paws, pawpads, petite,
Important info: See my showcase images. Her accessories and attire can now be fully adjusted. Remove the accessory tags and put them in the negative prompt. Some accessories such as the anklet and sandals can be a bit more finicky than the others, generate a few extra images and these should disappear. Despite heavy use of images that show her feet clearly, she does not always have the correct number of toes, it is what it is I'm afraid. Extra tags that are useful for her feet include: paws, pawpads. "Petite" is useful to get her closer to the V1 proportions, since the proportions differ somewhat in this version. When the anklet and sandals are being finicky and you have removed most of everything else, the tag "nude" can help.
Krystal, originally introduced in the GameCube game Star Fox Adventures, hails from the planet Cerinia. An unknown calamity led to the destruction of her home planet and the loss of her parents. Somehow surviving this catastrophe, she leaves on a search for answers to how and why this happened. A distress call eventually brings her to the Lylat System and Dinosaur Planet. After a series of unfortunate events, she is encapsulated in a crystal and eventually rescued by Fox McCloud. After the events that unfold on Dinosaur Planet, she joins the Star Fox team.
I mean, does she really need an introduction? She is our favorite furry vixen. Solely responsible for the creation of millions of furries all across the globe. She is the OG. Yes, there are tons of Loras for her already and she is baked into the majority of Checkpoints. This is my take on her, I try to remain somewhat faithful to the videogame look and proportions while giving the character my own spin.
Recommended checkpoint: https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl
V1 Trigger word: ChickenKrystal
V1 tags: furry female, two-tone fur, blue fur, white fur, blue hair, teal eyes, animal nose, fox ears, fox tail, sharp nails, clawed feet, leg tattoo, arm tattoo, loincloth, strapless bra, tail ornament, hairpiece, circlet, metal collar, chain necklace, shoulder armor, bracer, anklet, sandals,
Conditional tags: from side, from above, from behind, cleavage, tail raised,
Extra info: Use the conditional tags when depicting her from these angles. The tag "cleavage" was used in a specific image "from above" and "from side", but of course works anytime. Most of the ornaments can be removed by removing their tags and/or putting their tags in the negative prompt. Except for the headpiece, circlet, collar and necklace which appear to be quite stubborn. Play with the strength if you feel it doesn't listen to you. The bra and loincloth can be removed or replaced by something else by using other prompts for those regions. For example, she wears white panties in the game and those can replace the loincloth or be added under it if you like.
All images generated locally using the recommended Checkpoint at resolution 1152x1528. Images generated on CivitAI at 832x1216 will look similar, but not identical. Do not download the showcase images and extract the info using "PNG Info", as some info has been adjusted manually before uploading them to CivitAI. It will not look correct, enter the information manually in your generator.
Description
FAQ
Comments (20)
Another great job, not one of my favorites but Krystal is still a classic.
Bro even did the promotional posters that I remember staring at for far longer than I should have when I was like 14 or 15.
She was the 'final' of the four horsemen of my furry apocalypse that sealed the deal. Prior to her... Sally, Lola, Rena..
I like the stylistic Krystal lora's but always wanted a more game accurate one.
Thank you sir. Sadly, I don't have any buzz to tip you or I would.
Sadly your last 3-4 images don't have their metadata, since they were IMG2IMG.
Thank you! The overall theme of the Loras Ive made so far is that I want to have a somewhat stylized look, but also retain the proportions from the games. Every picture in the dataset for this one is based on additional IMG2IMG done on 3D renders from the actual model. With that said, the generation process will always do its own thing sometimes, for example her upper body is almost always a bit longer than Id like. But what can you do?
It was a rather complicated process to figure out, but I'm still new at this so there's a fair bit of trial and error involved. I started out with the promo material, upscaled it and Photoshopped the eyes to increase clarity. I downloaded a 3D model of her Smash Bros Ultimate assist trophy model because its much higher resolution than her Star Fox Adventures equivalent then posed it and took pictures from many angles and distances. I used a Lora on this website to do IMG2IMG on some of the renders, which gave me some extra material. This led to my first Lora, which I then used to again on the renders to further stylize the look. Using a combination of AI generated IMG2IMG on the 3D models in various poses, plus the 3D models and some of the promo material, I finally had a dataset I could use. Everything went through Photoshop to get better eyes, (and yet, they're not quite as clear as I'd like, but I feel like I've done all I can in that regard) increase saturation and to manually restore some of the tattoos on her arms and legs. (the tattoos on the 3D model were broken, so half my material actually lacks them, hence the faded look...)
At one point, I had a version that got very close to the 3D model itself, too close even. It looks too 3D. https://i.imgur.com/ztKZFRh.png
Anyways, Sally and Lola are both very nice! I am not familiar with Rena however. As for the IMG2IMG pictures, the prompt is nothing special. The one below is for the one of her laying on her stomach. ("the pose" is the name for her exact pose in that picture on Danbooru) Basically just every trigger word for her, then whatever is relevant for the picture. (The ones with the staff, for example, add "holding staff" and/or "smile" depending on her face in the original.
masterpiece, best quality, amazing quality,
BREAK, ChickenKrystal, furry female, two-tone fur, white fur, blue fur, blue hair, teal eyes, fox ears, fox tail, animal nose, sharp nails, clawed feet, strapless bra, loincloth, headpiece, circlet, metal collar, chain necklace, tail ornament, bracer, shoulder armor, anklet, sandals, smile, the pose,
BREAK, detailed eyes, detailed hair,
OhNoMyChickens Ah, thank you!
And Rena as in Renamon since I grew up around that era. I'll do some testing later since sadly I woke up so late today that it's basically already time for work today.
WorstAirtist Oh right right, Renamon is absolutely one of the top furries from way back when! She appears to be quite well represented on here, so I won't touch her. I think my next Lora may be Coco Bandicoot from the latest Crash Bandicoot game where she got a brand new look and nobody has done that look yet. Sorry, I can be a bit long-winded, but I rather enjoy this process and the making of Loras! Even now as I look at this one, I see issues that could be corrected, but for now its more than good enough.
OhNoMyChickens I don't mind long winded, lol. Just was short on time.
However I did notice something: Did you upload the correct version?
I ask because when I started testing after work, I kept getting gens that looked nothing like the example image, despite using the same Checkpoint, and changing nothing of the info (as I use PNG info to do 'lora tests' to see how close I get to the example images.
I tried other checkpoints, altering the prompts, etc, and then i realized something I forgot about. A long time ago I remembered that if the PNG info has a different version, that you either do not have, or do not have anymore as perhaps you replaced it... the image will come out as though the Lora is not there at all (because your lora prompt will reference something that does not exist!)
So I checked the Lora in the PNG info of your examples WITH metadata, and it shows up as:
<lora:ChickenKrystalV8:1>
(version 8)
When i manually took out the 'lora prompt' and then put the lora back in, it shows up as:
<lora:ChickenKrystalV5:1>
(version 5)
Though when downloaded it shows up as:
ChickenKrystal.safetensors
Once i manually replaced the lora prompt text (aka clicked the lora itself to place it into the prompt box) i got an gen that looked more or less like it should with your lora.
So I figured it out how to get images that look like yours! but I'm not sure if the download is supposed to be Version 5, or Version 8.
(also love Coco, so I look forward to it.)
Or for a summary of what I said above;
The example images uploaded here were made with a 'version 8' while the download here downloads 'version 5' and thus one has to manually edit the version number, or delete the entire lora prompt, then click the lora from SD's lora tab to place the correct version in when using PNG info as a base.
But after figuring this out, I WAS able to get images to generate that looked like they were supposed to with the lora.
WorstAirtist Phew, yeah, the prompt had the name of the Lora which I believe was something like "<ChickenKrystalV8:1>" or something like that, which I then manually edited to be just "ChickenKrystal" in the image info to match the name of the safetensor file I uploaded. I did re-download the file myself to test verify that I uploaded the correct one immediately after I published the Lora. So while you won't get an exact match with the example images and their info, it'll be very close. I assume that changing even just one letter, even if it might be the name of the tensor file, that has an effect on the seed.
I just tried it too and the image comes out very close, but not identical. However, it definitely is the correct file. It's a bit of a mess on my computer, I went through about 15 different versions of the file before settling on this one! So my Lora folder is full of Krystal files from V1 to V10, with 2-3 epochs for each.
https://i.imgur.com/5uWKxna.jpeg
WorstAirtist Here are the original parameters used for the first image, before I manually edited the name of the lora and the prompt info for the pic on CivitAI. (I didnt want my chosen file to be called ChickenKrystalV8-000010.safetensor or whatever it was that CivitAI picked)
parameters
masterpiece, best quality, amazing quality, outdoors, grass, flowers, palm trees, stone ruins,
BREAK, ChickenKrystal, furry female, two-tone fur, white fur, blue fur, blue hair, teal eyes, fox ears, fox tail, animal nose, sharp nails, strapless bra, loincloth, headpiece, circlet, metal collar, chain necklace, tail ornament, bracer, shoulder armor, smile, small breasts, head tilt, looking at viewer, upper body, portrait, hand in own hair,
BREAK, detailed eyes, detailed hair,
<lora:ChickenKrystalV8:1>
Negative prompt: bad quality, worst quality, worst detail, sketch, censor,
Steps: 25, Sampler: DPM++ 2M SDE, Schedule type: Karras, CFG scale: 4, Seed: 3403697435, Size: 1152x1528, Model hash: 681815bb31, Model: waiNSFWIllustrious_v140, Lora hashes: "ChickenKrystalV8: ceeec42ab889", Discard penultimate sigma: True, SGM noise multiplier: True, Version: f2.0.1v1.10.1-1.10.1, Module 1: sdxl_vae
WorstAirtist Alright, I went through my files and re-tested the showcase post files I have locally. So I think the name of the Lora itself has an effect on the seed and thus changes the outcome a little bit. If I download the file from here and simply call it "ChickenKrystal" then the result changes a little. But if I call it "ChickenKrystalV8", then the results are identical with the images in the showcase gallery. Weird that something like that would have an effect on the seed.
OhNoMyChickens Likely it will work on your end because you actually possess all the versions. You would have all the versions so, regardless of which one you used, it would work anyway. Unless you deleted all the versions, other than the one that was uploaded. (or at the very least still possess v5/v8)
Or to clarify.
Downloader Joe downloads, and only has download version, drops example with v8 into PNG and generates (and doesn't know that his version is v5) Gets a fox creature that is maybe 40% krystal, but otherwise looks nothing like example images (I.E. very dark blue fur, and different physique)
Creator of lora tests the Lora, and it works regardless because creator of Lora has all versions in their Lora folder, so regardless of which they use, it will work. (again, unless you removed/deleted all other versions.. in which case I have no explanation!)
But yeah, as you said, if you alter the Lora prompt in the box in any way that doesnt match with an existing lora in the lora folder, it will no longer call on the lora since the text would be incorrect.
In fact here, I'll gen 2 images and post em on the lora page. The first will be one where i make sure its correct, and then the second one will be what i was getting before I realized the version mismatch, should make it easier to see!
OhNoMyChickens Ok. I just uploaded 2 images to the gallery. The first one is the same as one of your example images but genned from my PC with the correct version, which is v5. (the one we download)
The second is the EXACT same image, but using the default version of the PNG info, where it still says V8
I figure showing is easier than telling in this case.
And to be clear, it's all fine. I'm sure many people don't do like I do where they I PNG info as a base, so they didn't run into it! So you haven't done anything wrong, I'm not complaining, etc! (or what I mean is a lot of people manually write in their own prompts, INSTEAD of using PNG info as a starter) And the ONLY reason I had any issue was because I did use PNG info.
And i resolved it once i remembered that having a different version in the PNG info than the version you actually have, causes what you see in the second image!
So you don't have to do anything!
Was mostly curious since there was a version difference, and I wasn' sure if v8 was supposed to be the one uploaded or not.
Or a TLDR;
I figured out the version mismatch, and am only showing the 2 examples in the gallery so you can see what i saw at first.
WorstAirtist It's so weird that it would call on something different, when it's all just text and such. So I guess it tries to use a Lora and it doesn't match somehow, so you get a picture that only uses the Checkpoint and prompt, but not the downloaded Lora. So as long as people download the Lora, then manually copy the prompt instead of doing "PNG Info" and then send to text2img, it's fine.
WorstAirtist Just checked your picture. Yes, that is what you get with the prompt and checkpoint only, happened to me too locally when I forgot to load the Lora. So it's all good!
OhNoMyChickens Yeah, like i said, you didn't do anything wrong (and i'm trying to make that clear since a lot of what i said might make you think you did!)
It's only an issue for people like me!
And the only reason i test like that is because i have ran into issues in the past. with a small handful of Lora's where even when everything was the same (even the lora) the example images, and the ones I generated looked very different despite all else being the same.
ironically enough, with a Lora I commissioned. The one who made the lora, his examples came out amazing. Mine came out.. acceptable.
Turns out he was using a very niche checkpoint to get it to be as crisp as it was. Wouldn't come out that good on any other.
Once i got the checkpoint, my images came out basically perfect.
But that's also why I test! Because Ive had a few that didn't work as expected (one was a tifa lockhart) and i've seen LoRA's vanish. So I test, and if there's something off, I try to figure out what it is while the LoRA is still there.
Hell some lora's ive had don't have the encoded prompts in the lora (the thing you click that has a wrench, where you can save notes, etc) and so i had to go to the lora page, copy the 'trigger prompts' into notes, etc, because otherwise the lora wouldn't be useable!
Regardless I'm rambling at this point.
Your lora works excellently and I look forward to your future ones. ESP if they're gonna be more characters I like!
WorstAirtist Oh no worries, I understood the problem and that it was really just an unfortunate thing, but no biggie. That's why I only test these with a single Checkpoint and stick with it, typically WAI-NSFW since I find it produces excellent, clean results. I'd love to use, say, NovaFurry for it's enhanced furry capabilities, but with a Lora loaded, pictures often if not always come up blurry. I do use other Checkpoints when doing IMG2IMG to "steer" the Lora towards a style I like, then I change back and feed the images into training and then I finally verify with WAI-NSFW that everything came out the way I wanted it. I'm about to submit Coco for training, so we will see if my method worked better this time. 4K renders using ripped Crash 4 model, trained a Lora, did IMG2IMG with NovaAnime on those, some photoshop work, back into training to hopefully get a Lora that falls somewhere between stylized and the game. She has tons and tons of outfits and very baggy clothes, so I think the anatomy is going to be a challenge.
OhNoMyChickens Anyways, I'll gen some nice Krystal soon. Just got some stuff irl to do first, and some responsibilities.
Well i posted 4 gens but only 3 seem to have actually uploaded, and its been hours but theyre still not showing up on your gallery. I'll wait a day, then try to reup them.
WorstAirtist Saw em not, well done! The Lora appears to be doing its thing.
WorstAirtist Im struggling too with this. I post 6 images to a new Lora, only 4 show up. I manually edit the post to re-add, the still dont show up despite being analyzed and everything. I make an additional post with the missing images plus a few more, it takes ages to show up at all. CivitAI can be weird sometimes. Anyways. I posted Crash first, he was easy to get right while Coco will take a few more days. My Lora is good, but Im going on a trip for 4 days so thatll have to wait.
OhNoMyChickens Enjoy your respite!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.