CivArchive
    Lum (Urusei Yatsura, Base SDXL) - v1.0
    NSFW
    Preview 7982038
    Preview 7982474
    Preview 7982173
    Preview 7981964
    Preview 7981994

    Beta version of my Lum Character lora. Current version needs the lora loaded at strength 1.25-2 to show up on some models.

    Release version now up! Twice the training, 80% better model!

    128 weight version now up! Much better at anime style generation, its a heavier, bigger, full float-32 LoRa. Not suited to realistic generation, it works great to create a 2.5d artstyle using a realistic checkpoint though! (32-bit, still trained in 8-bit :)

    This version is an experimental fully 8-bit trained LoRa. Training required 5.89GB Vram.

    Current model is trained to 4 epochs, 920 iterations 8 epochs, 1840 iterations. For the beta, with the strong settings, all character details are there and correct. The release model requires less emphasis, but may need some additional prompting to correct the details depending on the style or scene. It's more versatile, and you can more easily change elements of Lum's character, but the trade off is slightly less adherence to the character from the trigger without extra LoRa strength. Further training should allow the LoRa to have more effect with less emphasis.

    Her signature bikini is BAKED in on the beta. To override it, you will need to emphasize the alternate clothes, and those clothes will almost always have a tiger stripe pattern. Which is canon if you know the anime. Lum always always wears leopard print. (I will eventually release a version trained differently, but this is how I did it this time :) The release version does correct this, by loosening some character accuracy.)

    Both models stay up, as they ended up being quite different! v1.0 is far more versatile, but doesn't fully capture the original Rumiko Takahashi feel. The beta is very rigid, nearly overfit, but produces Lum very faithfully. Pick your poison, or load BOTH for extra fun! (Play with the strengths, I've gotten some pretty good results.)

    No NSFW images were used in training. SFW model, if a bikini clad 70's style gogo dancer is SFW... lol!

    Works best with anime models. Release version also works great with realistic models! Try out a Pixar style Lum! Trained to capture the generic character, not the style, so if you want her in Rumiko Takahashi's style, use retro artstyle, 80s anime style, or some variation thereof.

    Description

    8 Epochs, 1940 iterations.

    Overall Better in some ways, a slight bit worse in others. Lora strength of 1 is now enough, emphasis still required on trigger. Horns sometime fail to generate or generate too large or wrong color.

    FAQ

    LORA
    SDXL 1.0

    Details

    Downloads
    195
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/15/2024
    Updated
    4/22/2026
    Deleted
    -
    Trigger Words:
    lum

    Files

    lum-000008.safetensors

    Mirrors

    CivitAI (1 mirrors)

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.