A small test to practice FLUX training which turned out to be pretty good. This LoRA captures the style of those pictures in retro anime style generated by midjourney.
Recommended at 0.8-1.2 strength.
There are no specific trigger words, so try to describe the exact style of retro anime you want to go for with detail and it will do the rest. (See the model images for examples, though you do not have to stick exactly with those)
If you find it cool or anything, please consider leaving a like, comment or other reaction!
Donations are massively appreciated!
Update plans: Making the grain/halftone much more controllable; Extending the dataset to not just MJ images; further extending versatility
Also feel free to post any images you create with it below, cause I'm really curious what people can get out of this LoRA ;)
Description
Init release
FAQ
Comments (58)
Why is there so much grain in the images?
It's a part of the style, the half tone effect which I'm currently working on to make it triggerable.
This lora is incredible! Thanks for sharing your work.
Where to put this lore and how to use it?
You use it like any other, except it's for the FLUX model.
I would like to ask, how much picture training have you used to achieve this effect?
What do you mean? The training strength? The number of images? Or...?
@LizardMan245 The number of images
@3years_go_Mars 186
look so fckgood
This looks great but unfortunately does not work in Draw Things. Any chance of a ComfyUI version that would work for these other platforms? Thank you.
This lora looks like its entirely trained from https://www.instagram.com/24kenva/. Their style is cool though so I'm glad to see finally a decent LORA replication of their style
Almost, but not entirely. There's some extra in it to make the style cleaner, but those are also images from midjourney so they're pretty similar, making the style look this way. In the next update though I'm going to further extend the versatility and range of the model.
whoops I just double posted this same thing, my bad, but yeah, the images look extremely similar to kenva's work. funny enough, the artist hates AI and made several pieces of art dictating his hatred towards it lol. he/she is a bit of a luddite when it comes to image generation in general.
Hi, I really want to try this LoRA out but when I use it in Forge it does not appear to do anything. Is there any special setting I need to set in the "diffusion low bits" dropdown like some other Flux LoRAs I've seen?
what flux model are you using?
@RatGF Hello, I'm not OP, but I seems to have the same problem. My base model is flux1-dev-bnb-nf4-v2.safetensors (default from a pinokio.computer install with forge)
Well, setting the "diffusion in Low Bits" to "Automatic (fp16 LoRA) fixed the problem for me
@Bob59000 yes, on nf4 the lora needs to be run at a higher precision to see the effects. 'auto fp16 lora' will run the lora at a higher precision and makes it easier to see the effects, but if using multiple loras it will run slower. You can also try setting the Lora strength higher to 1.8-2
@RatGF flux1-dev-bnb-nf4.safetensors
@Bob59000 I have tried this and no luck :( I will try again
@cubby_ai hmm try using nf4 v2, flux1-dev-bnb-nf4-v2.safetensors · lllyasviel/flux1-dev-bnb-nf4 at main (huggingface.co)
@RatGF I got it to work!! I updated my forge and now it works with the dropdown on "Automatic (fp16 LoRA)" still using flux1-dev-bnb-nf4.safetensors!
@cubby_ai nice ^^
love this grain bitmap quality it adds to images
A very cool Lora to play with,Thanks!
I'm surprised you didn't mention the original artist from which you trained these images. I haven't seen your actual training images, but they definitely look like they're from kenva directly. Might want to attribute that that was your source, especially given that the artist hates AI in general lol.
Yeah so I included a good amount of images from kenva in the dataset, that's true, however that was for the purpose of training that manga-like texturing, which those images have. The other part of the dataset is also made with midjourney images of practically the same style, but not from the same artist, making it a blend. On a side note I do not understand why would kenva hate AI as their images are made with midjourney besides the 3d art of course, but anyways I hope this clears it up.
@LizardMan245 Kenva doesn't use AI, at least not for generating. I believe they may use it for added effects such as adding extra shaders, however he/she paints them from scratch on a real live canvas. This is what makes the halftone effect, because they are painting directly upon an actual canvas that has those halftone bumps that also influence the painting, as well as the random artifacts that appear.
While I don't know the artist specifically, and many of these types of discussions have deleted comments so viewers will miss context of the social media interactions amongst anonymous users, this thing does exist:
https://www.instagram.com/p/C7zQfijKdjN/?hl=en
@skyger I'm gonna have to disagree. After some digging while I prepared my dataset, I found that kenva's images are really just perfectly selected iterations of niji midjourney images in a specific retro anime style with added contrast, grain, damage and a halftone overlay in photoshop. That matches with their profile description I saw somewhere where they said that they use AI and photoshop. I have seen other artists as well who explicitly said that it's midjourney whose images almost have the exact same composition style. For example: https://www.instagram.com/p/C1Okd_MIUKy/?img_index=1 and https://x.com/24Kenva/status/1717776956874162181/photo/1, or maybe https://www.instagram.com/p/C5Kq8uMKnf6/ and https://www.instagram.com/p/C9GuQzXp41k/?img_index=2. Also there are several discussions in kenva's comments about the same matter and the evidence leads to that these are highly retouched images made with the generator.
Why should he mentioned about it ? Why exactly ? Did Kenya discovered painting ? digital painting ? Did kenya discovered curve lines or straight lines ? If so, Someone discovered all alphabet and when you express something, you have to give credit who discovered alphabet as well =3
hey can you make this lora for 1.5 model with same dataset, this lora is super cool, i want to create videos using this lora , it would be so pretty
1.5 or SDXL will be great!
Unfortunately sd1.5 is not capable of learning details on this level like Flux, but I shall give it a test sometime. Maybe it's not entirely impossible.
this is incredible. Works great. Fantastic job!
can you train 1.5 lora with same dataset
something tells me this one is addictive... ^_^
1GB lora? Damn thats a chunky boi :P
You must have been cooking this for quite a while :P
Either way. Good job and thanks a lot!!!!
i agree 1 gb too much for a lora.
@bydm44 Yeah I wonder what makes some loras so large and others so small..
@cathyleverman quality and it's intricacy. But hearing the community in my next lora I'll try reducing the size without loss in quality.
@LizardMan245 optimum quality with lower size always good. remember first sd 1.5 files? lora was not exist and everyone was downloading 2 gb files :D
cool
Can I use generated images by using this lora for commercial purposes?
The base model flux doesn't allow commercial, so I think you can't.
@LizardMan245 Black Forest Labs officially confirmed that "Generated outputs can be used for personal, scientific, and commercial purposes" on their page of Flux.1 [dev] model on huggingface
cool retro magazine print effect is added to generations which is fun..but the CROPS ..omg. Everything is cropped like we are back in 2020
One of my favorite, thank you for your work, it is amazing !
Cool but unfortunately it's not possible to sell images with this LoRa. Waiting for a new version where Selling images is permitted 🙏
Nah I don't really want to get involved in this selling thing and all that legal stuff, but I can't say I will not allow it in the future, maybe for the next model😉
can you please make this for xl also T_T pleaseeee
I have already made one for xl with more or less the same dataset, however pony as it turned out produces far less vibrant compositions compared to flux. BUT this illustrious model seems more promising, I'll see what I can do.
Could you please tell me which checkpoint you are using? It doesn't work for me for some reason!(((
flux1 dev
Why is it that the standing character I generate always has the top half of the head cut off at the top of the frame and the bottom half of the feet cut off at the bottom?
describe hair type and floor type. try a different checkpoint
I like the shaders on this one
Love this
Details
Files
RetroAnimeFluxV1.safetensors
Mirrors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
45-RetroAnimeFlux.safetensors
retro-anime-flux-style_721039_806265.safetensors
RetroAnimeFluxV1.safetensors
F1_Retro_Anime.safetensors
retro.safetensors
retro.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
retro.safetensors
RetroAnimeFluxV1.safetensors
style_retroanime.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
Retro Anime Flux v1.0.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
RetroAnimeFluxV1.safetensors
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.