■This is an experimental ground for krea2 LoRAs.
■If you notice pixel patterns in your generated images, try the VAE below. It removes the pixelation typical of the qwen_vae and makes the images much sharper.
https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image
■If the VAE looks a bit too saturated for your taste, try using this VAE merge node to blend it with the original qwen_vae to dial it in. A 0.5 merge should be just right.
https://civarchive.com/models/2487550/vae-merger-for-comfyui-anima
Description
■This is a style LoRA focused on the aesthetics of cool, retro anime.
It was trained on a dataset of 18,000 carefully selected images from what can be considered masterpieces of traditional cel animation. You can find the list of the included works below.
■This LoRA was trained at 1024px.
I trained this using booru-style tag captions.
The number of epochs and steps might seem low compared to the size of the dataset. However, since I trained it with a Rank of 64 and an effective batch size of 16, I believe the learning process actually progressed quite well.
Interestingly, when looking at the comparison images, it seems like the model picked up a few unknown concepts in addition to the overall style, which might provide some unexpected benefits.
■depending on your seed or prompt, the LoRA's style might feel too weak. If that happens, please try increasing the weight. You will likely find a good sweet spot somewhere around 1.1 to 1.5.
or, Since Krea2 is already capable of reproducing high-quality cel styles on its own, applying this LoRA at a lower weight for a hybrid look is also a great option.
■Versatility:
Because the dataset includes a wide range of genres—such as Sci-Fi, traditional Japanese style, and fantasy—this LoRA is highly flexible and can adapt to almost any genre of image you want to generate.
■Character Genders:
Although I personally only generate 1girl images, this is not a female-only LoRA. Anime naturally features male characters, and about half of the dataset consists of men, so generating male characters shouldn't be an issue.
■About the Files & Backups
To stabilize the style, this main LoRA is a merge of several different training steps.
I am also sharing the unmerged backup LoRAs that were used as ingredients for this merge.
■List of Works: (The trigger tag does not exist.)
Cowboy_Bebop_movie&TV
ghost_in_the_shell
Rurouni_Kenshin_OVA
Memories
patlabor_movie_1.2.3
Vampire_Hunter_D
The_Vision_of_Escaflowne_movie
X (CLAMP 1996) movie
Millennium_Actress
GUNDAM_The 08th MS Team_OVA
GUNDAM_0083: Stardust Memory_OVA
GUNDAM_W Endless Waltz_movie
■A Personal Note on the Anime
I truly believe the titles in this list represent the absolute pinnacle of traditional cel animation. The artwork in every single one is on par with fine art, and seeing it brought to life with such stunning animation is a truly rich visual experience. They are all fantastic masterpieces absolutely worth your time, so I highly recommend checking them out if you ever get the chance.
FAQ
Comments (8)
It's sad that the old LoRA for Flux 1 Dev still remains the King of the Hill.
https://civitai.red/models/721039/retro-anime-flux-style?modelVersionId=2077750
Still relevant: generate anything with any model and then do i2i in Flux. Because no similar LoRA has even come close in terms of quality. Unfortunately, this one hasn't either.
I looked at the LoRA, and the style seems quite distinctive. It might be a truly unique creation that only they are capable of making.
Your lora seems to be very good and has good quality and consistency for the specified style. I love it
Thanks for giving it a try!
Hey, buddy! How do you decide on the style of the artwork without using any keywords? That image you posted in the gallery has several different styles, and it would be cool to know what some of them are.
Can you help me?
Thanks for giving it a try!
Just a heads-up, this LoRA is meant to be a versatile cel anime style for lots of genres, so there aren't any specific trigger tags. You'll always get that core cel shading look, but the subtle style shifts really depend on your prompt and the seed.
If you hit a style you love, it's worth checking your prompt to see what caused it. Like, robot or military uniform might give you a Gundam feel, while katana might lean toward Rurouni Kenshin. Even hair color or "day/night" tags can change the vibe. It basically reads your tags and automatically matches them to the closest anime style. The perk is that it does the heavy lifting for you, even if you don't know exactly what style you want!
Also, Krea 2 is already super strong with anime styles (cute, cool, retro, etc.). If you figure out prompts that work well for your ideal style in the base model, you've got a solid foundation. Add my LoRA on top of that, and you get that sweet cel anime synergy.
By the way, do you think having series-specific trigger tags would be better? Some folks requested it. I'm on the fence about it, but it would definitely be convenient. Alternatively, if you want a dedicated LoRA for a specific show from the list, I'm down to make one.
@hjhf Thank you so much for your reply!
Now I understand how it works, and it’s really cool, I’ve already managed to create some really interesting looks with it! My challenge has been maintaining a certain consistency in the images, because, as you said, everything changes depending on the prompt.
You did an excellent job with this Lora; she’s one of a kind on Krea, capturing that classic look with such high quality! My question about the trigger tags is because your image has some really cool styles that I’d like to replicate, but I haven’t been able to yet because I don’t have access to the prompt xD
The second image in the first column, where the geisha is in the middle of the fire, is incredible, and I’d love to capture that art style! I think the tags might help me achieve that
@pablosodres290
Ah, I've uploaded that image to the gallery, so please check it out!
You might notice some contradictory tags in the prompt, but please don't worry about them. I use wildcards to add random tags, so that just happens sometimes. I think the ones you really need to look out for are the Japanese-style tags. These elements probably have the strongest influence:
nihongami, kanzashi, lipstick, makeup, red lips, oiran, thick eyebrows, kimono, japanese clothes, floral print, hair stick
Also, if the retro vibe feels a bit weak, increasing the strength (weight) might help.
I checked my dataset to see if there were any Oiran (courtesan) images, and there were actually some from Millennium Actress, so the model might have learned that concept from there. There are plenty of women wearing kimonos in Rurouni Kenshin as well, but I don't think there were any with the traditional white makeup.
I'm really glad you liked it—it makes creating this LoRA totally worth it! Thank you!
If you have any questions or want to know the prompts for my other images, please don't hesitate to ask. I'm not entirely sure about the reproducibility (whether others can get the exact same results on their end), so I want to support you as much as I can!