I was basically cold and wanted to use my computer as a secondary space heater. I had a ridiculously large dataset lying around so ... Now I have a warm room and a decent Lora.Around 400 pictures handpicked, stitched, cropped, uncensored, or otherwise enhanced. Really, if you can imagine a picture of Nemu, it's probably in the data set one way or another. Didn't spend as much time tagging this time, but I still quickly reviewed and made sure the main features and positions/camera angles were tagged.<ul><li>The main trigger is Manaka Nemu. I am seriously surprised how well it work on its own.</li><li>You can use either anime screenshot or visual novel CG depending on the style you are looking for. In addition to the general style, expect subtle differences in eye and hair color as well as character design.</li><li>Purple eyes, brown hair, blunt bangs for the most obvious facial features.</li><li>French braid, single hair bun, sidelocks, for the main hairstyle. Add messy hair for a disheveled look. Braided bun and Nape can also be used for side or rear views.</li><li>Hair down, long hair, for the alternative hairstyle.</li><li>Black serafuku, black shirt, black skirt, thighhighs, loafers for the main outfit (also tagged but mostly redundant: school uniform, green neckerchief, sailor collar, zettai ryouiki)</li><li>Labcoat / white dress give the corresponding outfit.</li><li>Bondage oufit, harness / white collar work, but don't always give reliable results. See my other Loras about that.</li><li>Of course NSFW tags. The model generally doesn't show nudity unless nude is prompted or obviously implied. Female pubic hair works. Don't forget to prompt visible body parts, otherwise they'll be cropped out.</li></ul>Tested on Prefect Illustrious XL v7.0 and Wai Illustrious SDXL 14.0Also added a Lite version, half the size, similar performance for the general look, but not quite as good with small details (collar, harness ... I posted comparisons).

Manaka Nemu (from Euphoria)

This is a lite version, with half the network dimension and alpha (32 and 16) and shorter training (5 epochs). Less details than the full version, but half the disk space usage.

Files with SHA256: 1f536895f116b501d1b7faf60bc1e3d41695672e022630724802740e74d59c92