Type | |
Stats | 52 56 142 |
Reviews | (9) |
Published | Sep 3, 2024 |
Base Model | |
Training | Steps: 259 Epochs: 5 |
Usage Tips | Clip Skip: 1 |
Trigger Words | w3dworm, worm, 3d, black eyebrows, black eyes, small pupils, limbless, armless, legless, no humans |
Hash | AutoV2 B739A98603 |
❌💲 Commercial use of this model is not allowed
This LoRA works terribly, and it's not just about arms.
I can't experiment with finetunes because my PC can't even run Flux Schnell. I can't disable on-site generation. If there's a finetune that somehow gives better results I'd like to see that.
ℹUSE HINTS
If the worms resemble generic/stock worms, try increasing weight.
🖼DATASET
All original images are property of Team17.
Dataset does NOT contain images of worms in profile view or from behind.
Trained on just 9 images because that's all I could find in this particular style. Well, there were a few more pictures, but they were too small and low-quality and they looked more "chibi". What do you expect from materials for a game that came in the early 2000's? WFUS renders would probably make a better model, but I'm surely not spending another 2000 Buzz for at least the next 2 weeks.
I took the images previously used for my Worms render mix model, added slight Surface Blur to some images to cover up the JPEG artifacts and edited the captions (still short Booru-like). I added "no humans" to the whole dataset because AI may give worms human features (regardless of the base model) and there's no Negative prompt field for Flux.
💬EPOCH SAMPLE IMAGES COMMENTARY
Just to be clear, "OP" in the samples refers to me.
A couple of early epoch results show human noses, but thankfully no ears. I'm still paranoid though.
There seem to be issues with eyes across different epochs, including this published final one. I wonder if more precise prompting can prevent this.
Flux seems to be more obedient with the number of hands, only 1 out of 15 results got my request wrong in this regard (if you include the heavily deformed hands).
I don't know why the results for the prompt mentioning another LoRA of mine shows only very short weird text or none at all. AI doesn't like bad words? (like the "sucks" which is childish in the given context) Or Flux is trained not to allow mean-spirited texts?
Edit: maybe Civitai generator just can't handle texts longer than short, I can't test this locally with my insufficient PC. In fact, the PC started freezing when I tried to load a checkpoint in Fooocus and update a massive videogame simultaneously, I heard strange sounds coming from it, that was creepy.