Lain Iwakura from Serial Experiments Lain, the best anime of all time in many regards.
08/08/2026 : All hail the Lain MiniMax lora ! Brings out amazing video, very canonical. Please share your creations here !
V3 - 12/11/2025 : I finally got the nerves to do my V3 dataset because I wanted badly to have anime Lain on Qwen. The dataset is approx 150 faces directly from the anime plus supporting pics for the school outfit, bear costume and green cyberspace background/blue static noise.
It works very well but sometimes seems to have minor line artefacts/glitches. I think they won't come out if you prompt stuff like "Professional anime" / "CGI anime" / "2.5D anime" and/or use negative prompting such as "artifacts, low quality, etc"
It's able to do the really 90s canonic style of the anime as well as really modern yet very accurate
IT WORKS VERY WELL WITH QWEN EDIT 2509 and if you bring in an anime Lain screenshot as a reference picture, the quality becomes EXCEPTIONAL
I intend to do later another Lora to extend this one with a lot of outfits and accessories. We'll see ! Meanwhile, have fun.
Let's all love Lain ''^_^y'
NB : Special credits to @Lain32_ for essential support and teamwork with this LoRa. Make sure to check her great Pony/Illustrious LoRas, which are just the finest a Serial Experiments Lain fan can download.
OLD comments below about V2 :
This one took 12 hours of training with 17250 steps and a 575 pics dataset (courtesy of Lain32_ which gets all the credit for carefully getting and selecting pictures from the anime, and making the awesome Pony version)
Trigger word : Lain Iwakura
This was not expected but this model is very good at making semi-realistic or even realistic pictures of Lain. I don't know yet exactly how to prompt them successfully, but you can have a look at my examples to see it can be done. Using a LLM such as chatGPT is essential to get the most out of this Lora (as for any flux generation really). Samples generated using Acorn's is spinning flux checkpoint but I haven't tried ressources more fine-tuned for anime.
Description
I've skipped training here on a few recent models, but MiniMax H3 is just insanely good !
Now you can make canon-like Lain videos and even more ''^_^y'
FAQ
Comments (4)
Lets all love Lain <3
Not familiar with the anime but in your experience does training the character also retrain the style well in both ref2v or t2v/i2v? I'm anxious to train characters but am afraid that it wont be worth it yet.
This is the first Minimax anime lora i'm training so I can't really say. But i'm really impressed by how close to the original anime is that video where she stands in the street. It's exactly the vibe. I used a chatgpt prompt to mimic the anime but still, the actual results make me feel I could do a trailer for the anime using minimax. It's really spot on. I don't see why you would be anxious anyway ? Just try it and you'll see. This Lain lora is incredible imho and I didn't even try to optimize anything in terms of training
Thanks for sharing. May I ask if your dataset consists of images or videos? The results I got using my Anima image training set were absolutely terrible.