Lain Iwakura (Yoshitoshi Abe) - FLUX - QWEN - CHROMA - Serial Experiments Lain
Details
Download Files
Model description
Yoshitoshi Abe is a mangaka with a unique style, and the official character designer of Serial Experiments Lain. While his drawings serve as reference for the TV Anime & Playstation game, the style in the TV Anime is still quite different from his works. The Playstation Lain however is much closer to his style.
This LoRa uses exclusively Abe's own Lain drawings in the dataset, which includes the Playstation Lain character (but nothing coming from the game itself)
No TV Lain trespassing here, sir.
QWEN Version
This is my very first Qwen Lora. It uses the exact same dataset and captions.
I don't know yet how similarly to Flux it can be prompted.
It seems to output sometimes more "sketchy" images because the dataset indeed has pictures where aBe drew with some sketchy parts, so it seems to be higher fidelity than Flux, which can be a good or a bad thing depending on what you want to achieve.
For now I like to add "Sharp Clean professional CGI anime illustration" to my prompts
CHROMA Version
Works ok but on my setup pictures come out a bit blurry. I don't know why since I've trained a few Chroma LoRas before with the same settings and had no issue. Trained on Chroma-base fp16.
I just thought since I had the dataset I could just train it on Chroma and then we'll see.
It's not like we got too many Chroma loras at the moment and Lain is supposed to be everywhere right ?
Sample pictures were refined using an upscaler
FLUX V1
It can do Playstation Lain very well, in fact if you don't prompt the outfit you might sometimes get some elements from PSX Lain
Flux is so accurate I think it caught even the "grain of the paper" / "grain of the ink"
Some pictures ressembling too much some Abe drawing can be a bit blurry. But it's not systematic.
A v2 could potentially be done after doing some denoising / sharpening on the (very high quality) samples pictures... But I'm just saying that as a notorious Lain perfectionist...
The dataset contains Lains in multiple hair color hues and lighting so you might want to specify what you want
Overall a strong and very pleasant Lora from what I see so far.
I've tried to include the most interesting outfits as keywords but there are more.
Will try to give examples in the description later on.







