Inkbunny

IB Gear Latest Popular Members Search
 
Welcome to Inkbunny...
Allowed ratings
To view member-only content, create an account. ( Hide )
Cute cub Scar
« older newer »
AutoSnep
AutoSnep's Gallery (100)

First LoRA: Kit Cloudkicker
Nalazebra, the African Catdog
set default image size: small | medium | wide
page 1
page 2
page 3
page 4
page 5
First in pool
Cream on a safe solo adventure
+3
First in pool
Last in pool
I wanted to generate images with characters from TaleSpin, but apparently no models know the characters. There's a Rebecca Cunningham LoRA on Civitai and that's it. No Kit, no Molly. 😟

I decided it's a good reason to try fine-tuning a custom model. The first model is for Kit.

There's nothing fancy about the generated images, they're just raw outputs from Furtastic 2.0 with my Kit Cloudkicker LoRA applied. The dataset is primarily screencaps from the TV show, which I upscaled with 2x_Sharp_Anime_v1 ESRGAN model and then cut to pieces.

Preparing the dataset and adjusting parameters, when you know nothing, is a royal pain in the butt. But I'm happy it's finally working after the whole day of creating it. 😎

Judging by discussions on the Furry Diffusion discord server, to stop the model from forcing the style from the dataset too hard (without a complex prompt, the images with Kit are always flat colored currently), people add generations from the first variation of the model. But that's for tomorrow.

Download Kit Cloukicker LoRA v0.2-alpha on Civitai.

"
prompt detailed fur, hyper realism, kitcloudkicker, green sweater, ((blue-red cap)), bottomless, featureless crotch, solo, male, young, cub, short tail, anthro, standing, smiling, highly detailed background, forest
negative-prompt flat colors, cel shading, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
sampler ddim steps 28 scale 7.0
seed 1768625633, 823777158, 1985249079, 1983917303, 1275623957
Keywords
male 1,120,300, cub 252,393, anthro 192,559, solo 96,582, young 59,573, bear 45,452, smile 27,645, bottomless 23,546, happy 17,659, forest 13,495, ai generated 12,151, standing 8,759, solo male 7,195, sweater 5,190, stable diffusion 3,543, detailed background 2,345, talespin 1,679, brown bear 1,475, automatic1111 1,337, short tail 1,084, featureless crotch 820, kit cloudkicker 748, baseball cap 507, furtasticv20 233, detailed fur 180, furtastic 17, green sweater 9, blue-red baseball cap 2
Details
Type: Picture/Pinup
Published: 9 months, 1 week ago
Rating: General
MD5 Hash for Page 1... Show Find Identical Posts [?]
Stats
2,027 views
105 favorites
58 comments
BBCode Tags Show [?]
 
58 Comments on this Submission
Leave a Comment...
624aSecret
624aSecret
9 months, 1 week ago
link
omg a LoRA of Kit Cloudkicker now we just need a lora of molly
AutoSnep
AutoSnep
9 months, 1 week ago
link
The images for Molly's dataset are ready. 😎
SkunkBoy
SkunkBoy
9 months, 1 week ago
link
Where is his scrumptious little cock and balls?
AutoSnep
AutoSnep
9 months, 1 week ago
link
Adding more elements to the picture reduces the quality of the result as I have to carefully balance weights between the LoRA and the style. Once I mix the generated images into a new version of LoRA, the model should be more reliable and should be able to handle the NSFW stuff without breaking down.

I've created NSFW versions, of course, but they aren't worth publishing yet. You'll have to wait a bit. 😉
Alexlavendel
Alexlavendel
9 months, 1 week ago
link
Edited 9 months, 1 week ago by Owner
If you would like I can walk you through enhanced inpainting. The idea is that you start with the best image you can get out of the tool, and then open it in photoshop, draw it in roughly, zoomed way in if that's what it takes for good cursor control, and then load that as an img2img for inpainting. ONLY inpaint over the part you wish to transform from scribbles to not scribbles, and turn the denoising strength to 33% or so. It can be less if you drew it well, more if it's badly drawn.

Holding ALT will flip you between brush and sampler, just sample the colors from the image.

I usually just export the edits right over the PNG I was editing, so it's easy to find in A1111.

And remember, it takes three colors to make an object "turn".
AutoSnep
AutoSnep
9 months, 1 week ago
link
What do you mean by "turn"?
Alexlavendel
Alexlavendel
9 months, 1 week ago
link
so, like, if you imagine a cylinder shape, and then separate by thirds vertically, almost any three different colors, or shades of a color, will make the surface "turn" along the interface of the colors, but it takes THREE colors to make it really pop. Because it's inpaint it does not have to be great, it just needs the three colors and maybe a bit of smearing it around with a smudge tool on low strength. It's probably the most important thing anyone ever actually taught me about this.

Usually I sample the colors from elsewhere on the image, like a highlight, a shadow, and a midtone. If you really want to get fancy, you can add just a little bit of whatever color is on that "side" of the piece or a bit of whatever color is right next to it on whatever side. its just like oil painting, just put in a dab in the area and smear it in. The more you use the more "reflective" it seems wherever you added the color, and causes "gloss" to happen if well distributed on the side.

I'm stoned and its late and I've been on an art bender touching up some pieces that I immediately decided I hated after posting them and sometimes I just ramble things I'm thinking about my own art as well (these are things I'm trying to learn to apply well now that I have a tool that makes different mistakes than I do.

Please don't mind me if I'm bothering you or you know this.
AutoSnep
AutoSnep
9 months, 1 week ago
link
No-no-no, don't worry. Even if you're saying something I already know, it's always easy to miss something obvious which is common knowledge for someone else. 😆

When I'm preparing an image for inpainting, I usually do something like you describe, just without the exact science of counting colors. And then usually just apply Gaussian blur until the image looks okay-ish from a distance.

I've had issues with one image in particular where SD refused to properly connect a tail to a crotch (it drew a huge fluffy tail for a lioness initially). I spent a lof of time trying to get the parameters and inputs right, but SD just refused to function in this particular case for some reason.

Now that ControlNet exists, are there any new fancy ways of doing inpainting? Like telling SD, "A tail goes into this thin area and nowhere else" or something like that. 😁
Alexlavendel
Alexlavendel
9 months, 1 week ago
link
Edited 9 months, 1 week ago by Owner
I WISH you could apply a controlnet with a textual description like "the tail on this" as a "source", and then inpaint where to put it. That would be dope. But no. So I just paint it all instead, and then inpaint... Then repaint with more color... Or occasionally use lasso, copy paste, and then distort it into shape using free transform and control-dragging the corners.

I also like to snip, increase image size on the snip, inpaint all but the borders, and then downsize and re-insert by erasing the interface border.

It really helps when the nose is small and it mangles the details...

Anyway, I'm mostly learning, myself even if I learned all this 15 years ago and never applied it since.
AutoSnep
AutoSnep
9 months, 1 week ago
link
Yeah, that's a lot of trial and error. Makes me wonder why people think video generation of decent quality is anywhere within reach. 😆
Alexlavendel
Alexlavendel
9 months, 1 week ago
link
Or why actual pieces with large degrees of added work don't have artistic merit.
AutoSnep
AutoSnep
9 months, 1 week ago
link
They have a point though. It's a different kind of work. Being a very picky commissioner who puts red lines over an image in MS Paint and demands dozens of edits with variations does require creativity, but it's a different skill set compared to an artist who uses a brush to draw the image. And in case of LoRAs, it resembles an engineering job way more than an artistic one.

In time, as more artists inevitably start relying on ML more heavily than now, and as the tools advance enough to give very precise control instead of rolling a dice, the line will get more blurred, but currently the distance between an artist and a director is still very significant.

I intentionally never use the words "AI artist" because for me, it'd be a sort of Guinea pig who isn't a pig and and isn't from Guinea. 😆
Alexlavendel
Alexlavendel
9 months, 1 week ago
link
Yeah, for me I use the term "artist" because when you dump 8 hours into redrawing every detail, and seeing in your head what you will demand out of the AI without accepting anything less, that IS art. Granted what I'm going for is actually LoRA production, not individual pieces.

Then, I understand your process and mine are different in their own ways as well and that's OK.

For me I actually make the edits instead of just redlining. Oftentimes I have to tell it to put in fur by just drawing fur, but shitty-like with flicks of bright and shade. The AI is just for cleaning up my sloppy work and suggesting details and concerns in a piece I would have overlooked.

Heck, sometimes I will even straight up composite outputs, like if I like "feet from here, face from here, ears from this one" I will either copy/paste-in-place/blend or use it as the reference forehand drawing and do subsequent inpainting.
AutoSnep
AutoSnep
9 months, 1 week ago
link
"Foot from here, arm from there" is called editing and I wouldn't call it artistic as it's mixing dice rolls. All of this is very far from a job you learn for 10,000 hours and then perfect for 10,000 more.

I'm not forcing anyone to use my definitions and standards. Some people call 5 year olds with a pencil artists just because they can hold a pencil and a piece of paper, and legally, they technically are (just like that gorilla who took a photo), but this doesn't fit my definition either. 😆
sankau
sankau
9 months, 1 week ago
link

It turned out beautiful! Congratulations!
AutoSnep
AutoSnep
9 months, 1 week ago
link
Yes, I quite like the result myself. I had my doubts though when the first attempt looked like this: https://i.imgur.com/64NGjWF.png 😜
SkunkBoy
SkunkBoy
9 months, 1 week ago
link
So you're planning a NSFW version?
AutoSnep
AutoSnep
9 months, 1 week ago
link
Of course. What's the point otherwise? 😁
FoxyIbLover
FoxyIbLover
9 months, 1 week ago
link
LOL
FoxyIbLover
FoxyIbLover
9 months, 1 week ago
link
Foxy: *throwing pants* "Argh, put on some pants, put on some pants!"
AutoSnep
AutoSnep
9 months, 1 week ago
link
Zaush won't agree with you, going bottomless is totally fine. 😁
FoxyIbLover
FoxyIbLover
9 months, 1 week ago
link
Foxy: *yelling* "He is a pervert! He is a pervert!"
AutoSnep
AutoSnep
9 months, 1 week ago
link
Are you calling all classic Disney artists perverts? I mean, maybe... 😜
FoxyIbLover
FoxyIbLover
9 months, 1 week ago
link
Foxy: *ranting* "All artists are perverts, they are all perverted! Their art is a blemish on our sacred reality!"
asthexiancal
asthexiancal
9 months, 1 week ago
link
Edited 9 months, 1 week ago by Owner
Wow, you did a truly EXCELLENT job, the only improvement I'd see to be made is on his feet (more claw-ish ends, perhaps only by changeing the color?) - that is, if you want to be 100% on-model, cause he is already fantastic the way he is and this part is relatively optional eventually.
Yes, I've been searching myself for Kit/Molly models (for 3D prints) and had the same problems as you on average...

Kit: an artist called Lemurfeature did a really great 3D model of him, but he sadly retired it from Artstation years ago where it was on display. He has a Twitter and I tried to contact him, but got no answer.

Molly: zero, nada~ BUT! There are at least two (I think) models of Sunni Gummi (The Gummi Bears), who differs really little from Molly as for the face, it should be extremely easy to modify one into the other.

Note that there are several Baloo models, very easy to find but half of them are really accurate, and again many creators retired their models - of course it was the best ones...
AutoSnep
AutoSnep
9 months, 1 week ago
link
Edited 9 months, 1 week ago by Owner
Feet aren't the only issue. The white collar is wrong, the sweater and the baseball cap are too random, the weird fluff on the crotch appears all the time (I mean, it won't be an issue for NSFW variations, but still). As for the feet, I did get claw-ish look in some of my generations. It's rare and inconsistent, but possible.

I'm currently working on improving the dataset. More feet, more back views, more outfits. I don't know how well LoRA can handle extra concepts like the cloud surfboard, but I want to try. I'm also going to try dropping (most of) the fanart part of the set as I suspect it doesn't improve the dataset's style variety significantly and introduces unnecessary anatomical randomness.

Hopefully, once the next version of the dataset is done, I'll be able to generate on-model Kits, even if with heavy inpainting. And with enough generations in different styles, the final version of LoRA can be trained with the new generations within the dataset. If I understood correctly, this how people get character LoRAs to be less fragile to complex prompts and not enforcing the original style as hard.

Looks like I chose the hard mode for my first LoRA. Oh well, that's one way to learn. 😁

I found this post by Lemurfeature. Is there anything else? I'm not that good at 3D editing, so I'm not going to try editing Sunni Gummi models into Molly. Seems like a risky path of introducing incorrect data. 😁

There're 2 models of Baloo on Civitai, but they seem to be a bit off. Not a major problem, but I think having a proper LoRA would be useful, so if Kit and Molly LoRAs work, I guess I'll do Baloo too. The existing Rebecca LoRA seems decent, but I haven't tried it.

Overall, if you have any images of the characters in non-flat styles, please share.
asthexiancal
asthexiancal
9 months, 1 week ago
link
Edited 9 months, 1 week ago by Owner
Of all the remarks you made, I only saw the belly fluff, but even if it's non-cannon I find it quite amusing and, well, interesting, I think it adds something... I also just noticed the tail, no main Talespin character got any (Kit, Molly, Rebecca, Baloo) but many others do indeed.

About Lemurfeature, I forgot he still got an example of his Kit model displayed on deviantart --> https://www.deviantart.com/daehanraney15/art/Kit-Cloudk...

As for the Talespin characters in volume, they are so rare - Baloo excepted once again, for some reason - I don't think I saw anything, let alone saved!

Can't wait to see what you're going to produce....................................   :-D
AutoSnep
AutoSnep
9 months, 1 week ago
link
I think I like them with tails more. 😁 The lack of tails in the screencaps makes the generated tails inconsistent, but whatever. 🙃

This post on DeviantArt is a low-res stolen version. It needs to be reported actually. Well, the image from Twitter would be enough, I guess, though if the model was available I'd take shots from more angles.
asthexiancal
asthexiancal
9 months, 1 week ago
link
Ah yes, you are right, I forgot that - this is how I came to know that work, but also how I saw its author withdrew it from everywhere else.
I believe this post has been already reported several times giving some comments, but it looks like DA admins don't care a damn. How surprising.
ThumpyBunny
ThumpyBunny
9 months, 1 week ago
link
Holy shit! You really captured his cartoony look, but in a realistic style.
I'd love to see more of him without his clothes on!
AutoSnep
AutoSnep
9 months, 1 week ago
link
If I don't finish my work on improving the Kit LoRA soon enough I'll post NSFW generations done with this alpha version. The results are a bit messy and inconsistent, but it generated a few nice dicks. 😁
ThumpyBunny
ThumpyBunny
9 months, 1 week ago
link
Nice dicks are always welcome. Can't wait to see the end result!
AutoSnep
AutoSnep
9 months, 1 week ago
link
Someone posted a NSFW image they generated using my LoRA. That director is into dicks with foreskin though, not the kind of dicks I had in mind. 😁

I'm continuing watching TaleSpin at the moment, growing my datasets.
ThumpyBunny
ThumpyBunny
9 months, 1 week ago
link
I believe I saw that one in your favorites.
This is all so new to me. I might try getting into generating some images myself once I figure out what skills are needed.
Sounds like you need to put a lot of hours into creating these models and training the ai, which I have no idea how to do, or where to start, haha!
I'm just glad you're willing to do all of that and generate some cool stuff, so I don't have to!
Teddy
Teddy
9 months, 1 week ago
link
Shame training a LoRA only seems to work well with one character.  I've dabbled as well because I wanted a thumbsucking LoRA.
AutoSnep
AutoSnep
9 months, 1 week ago
link
I've seen people use ControlNet to put multiple non-interacting characters into the picture. I don't know whether it'll work, but I'm at least going to try to put the characters into a picture, mix with z-buffers and then inpaint the hell out of all interacting parts. 🤓

I've also seen multi-character LoRAs (for Paw Patrol, for example). I don't know how usable this is, but the author of the model did get at least one picture with proper feral fuck between two characters, which is a good sign. 😄
Thunderstick
Thunderstick
9 months, 1 week ago
link
What fun! Do you plan to post this Lora on CivetAI or is it available anywhere now?
Thunderstick
Thunderstick
9 months, 1 week ago
link
Oh never mind I found it. lol
AutoSnep
AutoSnep
9 months, 1 week ago
link
Haha, nobody reads descriptions, it's normal. 😁
Thunderstick
Thunderstick
9 months, 1 week ago
link
Guilty! But I do wonder how hard it was to create a lora? Any tips for those of us who want to do some ourselves?
AutoSnep
AutoSnep
9 months, 1 week ago
link
Technically, I haven't even finished creating one LoRA. So I'm not the right person to ask about instructions. 😆

I think usually it's infinitely easier than what I'm doing. For more popular characters, there're tons of art in various styles. I, on the other hand, have to deal with primarily screencaps from the TV show, with only poor DVD quality available.

If you pick, say, Renamon, this problem just doesn't exist. You can download 10,000 images from any booru and use the included tags, with some editing and cutting.
nelson88
nelson88
9 months, 1 week ago
link
Well...this is a pristine piece of art!
Kit has been always one of my favs ever!
Each pic is just fantastic,love to see the mature version too!
This masterpiece will going to my favs now!^^
AutoSnep
AutoSnep
9 months, 1 week ago
link
I'm glad you like it!

Though it does feel weird that 3 out of 4 submissions using this LoRA aren't mine. 😆
nelson88
nelson88
9 months, 1 week ago
link
Hope to see more stuff like this soon!😁😉
AutoSnep
AutoSnep
9 months, 1 week ago
link
I'm currently stuck at the watching the show stage. 🤣 I realized I've never watched it thoroughly and that it's pretty good. Along the way, I collect screencaps with all major characters in different outfits, at different angles etc.
nelson88
nelson88
9 months, 1 week ago
link
That's nice!
My friend and I do the same when we work in our pics!
And about your talent and stuff...I think you are a pretty awesome in this field!Have fun!^^
sankau
sankau
9 months, 1 week ago
link
What program do you use to make these drawings? Indicate me any tutorial on youtube?
AutoSnep
AutoSnep
9 months, 1 week ago
link
All you need is to install Stable Diffusion (I use A1111 web UI) and download furry models (I get them on Civitai.com). I didn't use YouTube tutorials for installing and downloading, so can't point to any. You can just google for that, I'm sure you'll find something.
BlazeHeartPanther
BlazeHeartPanther
9 months, 1 week ago
link
Oh, if you've done Kit, you MUST do Molly.
AutoSnep
AutoSnep
9 months, 1 week ago
link
For what it's worth, the dataset for all major TaleSpin characters is growing. 😎
BlazeHeartPanther
BlazeHeartPanther
9 months, 1 week ago
link
Lets hope she shows up in the future.
TheDeviousOne
TheDeviousOne
9 months ago
link
Molly will be amazing, such underappreciated character she is.
Personally i would also love to see Sunni Gummi, i don't suppose there are any plans for her too?
AutoSnep
AutoSnep
9 months ago
link
I've never watched Gummi Bears and I doubt I'll watch the show now, without the nostalgia drive. So probably no Sunni Gummi. 😆 But once I get to Molly LoRA, you can try aging the character up or something.
sankau
sankau
7 months, 2 weeks ago
link
Could you be so kind as to upload the Cloudkicker Kit's LORA again? Thanks.
AutoSnep
AutoSnep
7 months, 2 weeks ago
link
sankau
sankau
7 months, 2 weeks ago
link
Thank you very much! I can't find anything on civitai. I need a more efficient translator. 🤣🤣🤣
AutoSnep
AutoSnep
7 months, 2 weeks ago
link
I uploaded it 5 hours 15 minutes ago. 😁
TheAtomicDog
TheAtomicDog
7 months, 1 week ago
link
Looks truly like Kit doing some costume tests during pre-production.
New Comment:
Move reply box to top ↑
Log in or create an account to comment.
 

Help

Information

Policy & Legal

Twitter Facebook LinkedIn

All artwork and other content is copyright its respective owners.

Powered by Harmony 'Gravitation' Release 80.

Content Server: Virginia Cache - provided by Inkbunny Donors. Background: Blank Gray.

The Inkbunny web application, artwork, name and logo are copyright and trademark of their respective owners.