Type | |
Stats | 175 |
Reviews | (15) |
Published | Feb 18, 2024 |
Base Model | |
Training | Steps: 7,960 Epochs: 8 |
Usage Tips | Clip Skip: 1 Strength: 0.8 |
Trigger Words | riccco99 |
Hash | AutoV2 67559D8F30 |
Trained on around 70 images from instagram accounts @riccco99 (the model's account) and @ojitoole (the photographer's account).
Not meant for commercial or NSFW use.
I made this LoRa mostly to learn the process of making and using a model better. It helped that these people have been producing some high quality photography.
The model is based on the AbsoluteReality v1.81 checkpoint.
The activating tag is "riccco99". (I accidentally used "riccco9" in some of the examples and it gave me excactly the same results, apparently SD treats these the same.)
This person has a lot of different styles of appearance in their photos. In this model, I didn't reproduce all of them, but you can at least switch between 2 main kinds of styles by prompting either for "short hair, black hair" or "long hair, brown hair".
The model works with anime checkpoints to some extent (not well with 2d though).
As a little note about the tagging, the auto-tagger that I used produced tags for the point of view as "from XXX", for example "from side" or "from above" so tagging like that should work better compared to "side view" etc.
Known Issues:
Eyes, especially "squinting" and awkward pupil position. I spent a lot of time adjusting the training set in order to reduce this issue, and it mainly still comes up in side views. This is heavily present in the original photos and so I couldn't seem to be able to get rid of it completely. It can also be thought of as a special trait of this character, if you are generous.
Collarbones being too prominent. I didn't initially detect this issue, and wasn't motivated to rework the training set again, so it's just going to be present in some of the images. Again, also a feature in many instagram photos, although the model sometimes takes this too far.
Being a bit weak with anything that's not a front view. "from side" and "from above" views still have a fairly decent success rate, due to having some support in the training set, but the model is not suitable for back views.