santa hat
deerdeer nosedeer glow
Sign In

GekiDol ゲキドル - LoRA Pivotal Tuning || Airi / Seria / Izumi / Doll / Akira / Manami / Mayuri / Kazeharu / Tomoko / Kaworu / ...

75
537
13
Verified:
SafeTensor
Type
LoRA
Stats
537
Reviews
Published
Oct 12, 2023
Base Model
Other
Training
Steps: 30,000
Usage Tips
Clip Skip: 2
Trigger Words
ADol-30000 Airi-30000 Akira-30000 Alice-30000 Hirokazu-30000 Izumi-30000 Kaworu-30000 Kazeharu-30000 Makoto-30000 Manami-30000 Mayuri-30000 Seria-30000 Tomoko-30000 xclmsA-30000 xclsB-30000 screenshots-30000
Training Images
Download
Hash
AutoV2
156B4D2970
default creator card background decoration
Holiday 2023: 5 lights
alea31415's Avatar
alea31415

This model can only be used with both embeddings and LoRAs so you must also download the embeddings from what I mark as "training data"

Using only the LoRA is not effective (civitai however does not allow to upload zip directly for LoRA. I will eventually upload the bundle format once it is supported in webui). The zip containing both lora and embeddings can otherwise be downloaded here https://huggingface.co/alea31415/YuriDiffusion/blob/main/gekidol/gekidol-civitai.zip


This model uses pivotal tuning as explained in https://civitai.com/articles/2494/making-better-loras-with-pivotal-tuning

The data construction workflow follows https://civitai.com/articles/2383/a-99percent-automatized-pipeline-to-construct-character-pack-training-set-from-anime

All intermediate checkpoints and pts available at: https://huggingface.co/alea31415/YuriDiffusion/tree/main/gekidol


Pivotal tuning trains embeddings and the network in the same time. For this model the content of the zip file is (no more LoRA in the zip downloaded here)

You need to put the LoRA in LoRA folder, pts in embedding folders, and the txt file can be used with https://github.com/adieyal/sd-dynamic-prompts and https://github.com/DominikDoom/a1111-sd-webui-tagcomplete for easier prompting.

This is a temporary solution. The bundle system will allow much easier usage of results obtained from pivotal tuning. See https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568

About the model

Honestly, I do not particularly expect people to use the model.

  1. Gekidol is a very unpopular anime, but a good one (well ... I suppose). I take this opportunity to do some promotion for it.

  2. The model is trained on https://huggingface.co/Crosstyan/BPModel. Normally this makes it work less well on popular models although this is less the case here thanks to pivotal tuning. BPModel has its benefit, for those who know what it is for 😉. (And yes, as what I usually do, you can still use other LoRAs or build merge on top of it.)

  3. As explained, the current ecosystem is not friendly enough for this kind of setup.

If you happen to want to use the model, here are some important associations

  • Seria: yellow hairband

  • Mayuri: hairband

  • Manami: glasses

  • Izumi: hair scrunchie (and as usual the model cannot tell left from right)

---

I plan to completely shift to pivotal tuning once the bundle system is incorporated and LyCORIS is supported in HCP-Diffusion. As you can see, this model has a quite strong anime styles, which for me is related to the use of LoRA instead of LoHa here.