Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy 😁
Wield a rainbow's radiance, bright and flat,
Create Colorful Anime Treasure with CAT!
The age of Pony is over, the time of the CAT has come! -- Lord of the Rings: The Return of the King (probably)
Update CAT IL v10 - (CAT_NAI_v2:0.6) x (IL_v2:0.4)
Update CAT IL v9 - (CAT_NAI_v2:1.0) x (AolaStyle:0.5)
Update CAT IL v8 - (CAT_NAI_v2:1.0) x (EnergyCAT:0.3)
Update CAT IL v7 - (CAT_NAI_v2:1.0) x (Unreleased_CosmicCuteStyle_Lora:0.6)
Update CAT IL v5 - (CAT_Illustrious_v4:0.5) x ((NAI x IL):0.5)
Update NAI v2 - This is (CAT_Illustrious_v4:0.5) x (NAI:0.5) so that its half and half. This way its closer to NAI's flexibility and knowledge than v1 but still maintains the high quality style of CAT.
Update NAI v1 - This is v4 merged with NAI in order to gain some of the Natural Language prompting abilities of NAI. This version should be compatible with NoobAI, Illustrious, and even a few Pony Loras. (There are several Pony Loras used in my showcase).
(rejected) Update v5 - Is more flexible with other styles and cleaner but loses the style of v4 so I'm going keep v4 as onsite generation. Here is a comparison of v5 and my baseline: https://civarchive.com/images/64724098
Description
FAQ
Comments (32)
Having trouble getting this to work. Using pretty much the exact same prompts as previous versions and I’m getting saturated/really dark images on the civitai generator. Older versions work just fine
Sorry to hear that! When you say older versions worked, which ones do you mean? Did NAI v2 give you better results?
IL v5 is this mix: (CAT_Illustrious_v4:0.5) × ((NAI × IL):0.5) --> Its only half of what v4 was.
It includes both extra Illustrious and NoobAI data, which makes it more—and sometimes less—sensitive to certain tags or words.
I’m not fully happy with the results myself, so I’m planning to release an updated version in the next few days.
Until then can you try remixing an image that you see was made on site and using that as a base?
Please let me know your findings so I can use that info to improve future models.
Sad to see the yellow-fication of ChatGPT images seep into your models, love the previous versions!
Oh really? lol my bad. In that case, I can stop mixing ChatGPT images into my Checkpoint.
I was intentionally trying to capture that style but I can simply stop trying for future versions. I have lots of other fun styles to mix into my checkpoint like Kawaii Mirai or maybe anime screenshot styles.
Can you please let me know if its bad in all the showcase images or only some? I know the "We are back" Megumin one definitely has the ChatGPT yellow glow but how about the others?
Thanks for the feedback hope to hear more about where I should take future versions. 😁
@CitronLegacy You don't have to change for little ole me, I'm sure some people love it, it's just become a bit of a nitpick that I can't unsee. It's more than just this checkpoint but many of your newer LoRa models as well. It's understandable though as I know how difficult dataset collection can be. Also I'm sorry, I appreciate all the work you do and hate to say anything negative about other creators.
Um, so that tinge is a ChatGPT thing? I confess I had no idea, I hadn't noticed it but now yeah I think I can spot it in recent citronL models.
My point being: 1- n=2 that it exists, 2- it's not very noticeable to an amateur like me, extrapolate to others, 3- my two cents is that I kind of like it. Not that I definitely prefer it, but maybe. I'd guess that makes it cpme down to individual tastes tho.
I'm NOT weighing in on the actual question of "should you keep it". I'm an amateur, with no model training chops. I'm just trying to give audience context.
(Sorry for the wall of text!)
@cranpeach
No worries — I really appreciate the honest feedback! 😄
I don’t see this as negative at all. I want to make art that’s fun and interesting, so feedback like this helps me course correct when needed.
So far, it’s 1 vote against and 1 vote for using the ChatGPT style — someone else actually reached out asking me to capture it, which is why I leaned into it for this checkpoint.
That said, I’ve definitely been intentionally channeling the ChatGPT look lately, so if I ease off a bit, I can probably get more neutral results.
Some of my recent LoRAs like "We Are So Back" and "Sauron’s Rave Tower" are 100% ChatGPT.
On that note — what did you think of the Kawaii Mirai style? Do you prefer that over the ChatGPT look?
@ravemry9
ChatGPT has a pretty distinct anime style by default. You can prompt it to shift styles (like saying “anime screenshot from a Ufotable show like Fate/Stay Night”), but it still tends to snap back to its own look. If you upload an image and say “copy this style,” it sometimes adapts — but not always reliably.
On the topic of individuals' tastes yes. I usually dislike the most common/popular styles, so I just try a bit of everything until something clicks. Since I can’t predict what people will love, I keep experimenting!
@CitronLegacy Glad to hear that, the world could use more of that! With the Kawaii Mirai style, I like the overall pastel color palette/style, the first image is great, the second image is what I don't like, if that makes sense, like where the white balance is off, but again some may like that. I can clearly see the ChatGPT style in it, but that could just be your checkpoint doing that if you're saying this Kawaii Mirai style doesnt have any ChatGPT images in its dataset
EDIT: If you have photoshop you can do something quick like this, doesn't work on all images but might!
@cranpeach Kawaii Mirai is also ChatGPT images but I was asking so I could get a better idea of how far the style bleeds into images. For all the images in the Kawaii Mirai dataset I prompted ChatGPT to give a specific style but the effect still crept into the lora.
CAT IL v7 is basically the ChatGPT version of CAT but I'll just go back to CAT NAI v2 as the neutral base (because that is IMO my best checkpoint) for my next mix.
Unfortunately that CatBox link didnt load for me. What did you do in it: adjust the hue and saturation of an image?
I don't have photoshop, but I can write a python program that will adjust hue/saturation on all images in a dataset as long as I know what adjustment is good. Although sometimes when I try to manually adjust hue/saturation/etc in a photo editor I never get the correct colors.
Hmmmm well this is an interesting problem to solve! It would be nice to use the power of ChatGPT in my datasets but I need to purify the images.
@CitronLegacy Huh it works for me, basically it just shows selecting the Curves tool, selecting the white eyedropper, and clicking the part of the image that should be white and hitting OK. Works on Photopea too. Here's a imgur link if thats better
I was going to say it is a little more yellow than I vs would like if you were counting votes @CitronLegacy
New Yellow vs Cat4
@cranpeach Yes that link worked! OH! I didn't know there was a tool that auto levels based on what should be white! Hmmmm I wonder if I can automate this.
Also: I like my photo editors the same way I like my women: with curves!
@CitronLegacy It's about even Between NoobAI and Illv4. I've been using Cat4 for longer and it didn't seem better enough to switch. I pretty much avoided using pony too much as I liked other SDXL checkpoints, but when I saw some Illustrious stuff I switched/added it pretty quick.
I just dont have that attachment to NoobAI
@cranpeach New version v8 has no ChatGPT images in it! I'd say its much more neutral than v7. Please let me know what you think?
(The Megumin cover image was made with a style Lora that was trained on ChatGPT art so ignore that one for your assessment.)
Speaking of that — for the style Lora trained with ChatGPT images (Buzz Me) — I used your technique and applied the white-eyedropper to all the images before training. The color correction was really nice!
@CitronLegacy Looks awesome! I'll have to test it out. Glad to hear the eyedropper trick worked out also!
@abigfavour That makes sense they are fairly similar, although NAI v2 works better with other styles.
Illustrious is easily 10x better than Pony but NoobAI is functionally the same as Illustrious. The main benefit of NoobAI is that it has much more knowledge than Illustrious so it adheres to prompts better and knows more characters/concepts/etc. You aren't missing much by staying with Illustrious (for now).
I hope this version is more stable then the last one, took me hours just to get three pictures out.
Hours for only a few images sucks.😨 Do you mean the checkpoint made bad pictures or that it didn't make pictures at all?
Is there something I can improve or was it a CivitAI problem?
@CitronLegacy No the pictures were good, it just took a while for them to render or even show up at all. I will say the new one works much better and is the slightest bit faster.
Just wanted to say, I love your particular style checkpoints. It says a lot about them that even tho I'm less of an anime guy myself, I often find myself coming back to CAT.
(I know this is the CAT discussion page, but to anyone reading, check out this creator's Galena chckpnt -- it's my fave Ilust. chckpnt. And I have excellent taste.)
About the model auction recs, I have no loras of my own (despite your encouragement, thx!), but can I suggest the Riju semi-real? https://civitai.com/models/1258967
Sounds like I'm joking, but I'm not, i love it and would love to see what more people make with it if it gets featured.
This is kinda breaking the (unwriten) rules since it's yours, so feel free to prioritize others' recs; I get it.
Also its totally ok for off topic conversation on this page. Its the central hub of my CivitAI existence so its a pretty good forum for any AI topic.
Glad I got you coming back for more despite it being anime 🤣
Theoretically these checkpoints will work with non-anime but if I advertise them the with non-anime they wont get usage lol.
Glad you like Galena CAT so much. Galena and CAT base are the two styles I've been using/refining for basically my entire career on CivitAI.
Lol well if no one else claims the free bids/buzz then yeah I'll bid on that Riju. Apparently I set the requirements too high because only one person claimed the bid so far.
Is this more NSFW than the WAI-NSFW-illustrious-SDXL checkpoint? if that makes sense
Ummm probably? I built CAT NAI v2 to be more flexible and knowledgeable than WAI. (I'm not claiming its better just more knowledgeable) Its 50% Illustrious and 50% NoobAI so its compatible with both Illustrious and NoobAI but it has all the extra knowledge than NoobAI has that Illustrious checkpoints lack. I'm going to assume that its knowledge includes NSFW that Illustrious doesn't know about because NoobAI's dataset is much larger than Illustrious v0.1.
Thank you for your checkpoints and other things. you put so much effort in these <3
NAIv2 Version of your checkpoints is a good one :)
https://civitai.com/models/1663431/spectral-dragon-homm5-nai
Thank you I really appreciate the feedback. 😁 Glad you like my works.
Ok good to know. I also prefer NAI v2 👍
Ooooh thats an awesome Lora! I bid on it!
@CitronLegacy thank you
I did some tests and CAT seems to be better than WAI-NSFW in anime but not cartoon/realistic stuff I think
Interesting! Thank you for experimenting! I'm surprised it doesn't handle cartoon well thats something I need to look into.
It makes sense that its not very good at realistic because I made this to be anime focused but making a realistic one someday would be nice.
Very impressive model, almost no need for lora!
Thank you! I reviewed a bunch of your loras. You have a great collection! 😁