CivArchive
    EtudeFT - v1.0
    NSFW
    Preview 58396399
    Preview 58396773
    Preview 58396811
    Preview 58396565
    Preview 58398086
    Preview 58398111

    This is my first trained model. It's a little late to say this, but it's very difficult.

    I used about 5,000 images based on v1-5-pruned-emaonly.safetensors and trained it for 100 epochs in 4 batches.

    The quality was very low as it was, so I mixed in 15% RM01 to stabilize it and added my own LoRA group to finish it. I think I could make it look like usual if I tweaked it, but there would be no point in using the trained model, so I'm keeping it to a light correction.

    Use Euler a for the sampler. Step 20 and CFG Scale 6 are good guidelines.

    Since it seems to be overfitting, the problem can be reduced by using the Hires Fix or the detailer.

    The body structure is still a bit iffy, but the fingers are easier to control than if I left it to someone else. This is partly due to LoRA, but I'm also starting to notice some burn marks, so it might be better to set the learning rate a little lower.

    Next time I'll have to increase the number of images by another digit.

    私の最初のTrainedモデルです。なにをいまさらという感じではありますが,とても難しいですね。

    v1-5-pruned-emaonly.safetensorsをベースに5000枚ほどの画像を使い,4バッチで100 epochsの学習を行いました。

    そのままだと非常にクオリティが低いので,RM01を15%混ぜて安定させ,自作のLoRA群を加えて仕上げました。いじればいつもくらいの感じにもできそうですが,Trained モデルを使う意味がなくなるので軽めの補正に留めています。

    サンプラーはEuler aを使ってください。Step 20,CFG Scale 6くらいが目安です。崩れ始めるのは顔からですので,Hires Fixやadetailerで対応すれば問題は抑えられます。

    身体構造はまだ怪しいですが,手指などは他人任せよりはコントロールしやすい感じです。LoRAのせいもありますが焼けも出てきていますので,学習率はもう少し低めでやったほうがいいのかもしれません。

    次は画像数をもう1桁上げないとダメですね。

    Description

    FAQ

    Comments (2)

    littlefluffyballFeb 22, 2025
    CivitAI

    Basically the output is the same as RMB01
    But this time it not drawing golden patterns and armors on everything.

    aueki4g467
    Author
    Feb 25, 2025

    The basic parts of the model are completely different from RMB01, but the finish is similar, so the output looks similar. Specifically, it is the part that corresponds to FinishingR. The adjustments in RMB01 and FinishingR are not the same, but the tendency is similar. I'm not sure why RMB01 was gold, but it was probably due to the compatibility of the differential learning problem.

    I just started FineTuning, so it's not good yet, but I feel that it is quite difficult to create a high-quality model with FT alone. If I adjust it so that it can produce a decent image to a certain extent, the tendency will be similar to the merge model. It will probably look the same for a while.

    Please wait until I can produce a solid image with FT.

    Checkpoint
    SD 1.5

    Details

    Downloads
    241
    Platform
    CivitAI
    Platform Status
    Available
    Created
    2/17/2025
    Updated
    8/29/2026
    Deleted
    -