This LORA produces a variety of inflatable toys, largely in the style of Imaginary Creature Factory products. It trained on three hundred photos, which I hand-labeled; style is therefore usually photographic but at lower weights the underlying model's original style can show up.
Prominent tags: ICFInflatable, inflatable Eastern dragon, inflatable gryphon, inflatable sea dragon, inflatable Western dragon, glossy vinyl, matte vinyl, etc. Has some knowledge of "partially deflated [limb]", "deflated", "flat", various colors of vinyl, various body parts, various poses, etc. Also has opinions about what various rooms and furnishings look like because I wanted to train them in so I could negatively tag them. Aware of vague "medium shot", "wide shot", "close up", "close shot", "front view", "front angle", "side angle", "side view", "back view", "from behind", etc.
Less-trained tags: a variety of names of specific toys, fox, ghost dragon, bat dragon, husky, maned wolf, convention, stacked toys, kissing, snuggling, tailbutt, paw focus, toe beans, inflatable talons, plastic talons (useful for inflatable gryphons), PuffyPaws inflatable, IWInflatable
Tags you may want to negate: IKEA, cluttered shelves, clutter, electric fan, etc. or put them in positively if that's the vibe you want. cluttered spaces are common in these photos (including the ones in my living room lul) and the model can overproduce that.
It's my first LoRA, I may try to improve the tags and add more content and balance it a bit, the usual. I wrote my own software to tag it, I have this entire taxonomy of tags now so I can just type "inflatable eastern dragon" and get all the inflatable eastern dragon parts you'd expect and just cancel out the ones that you can't see in the picture (and then they're marked so auto-deduce won't put them back in), I should finish up Just One More Feature and put it on github or something
Description
I've had much more trouble making the Pony version get decent results. I likely haven't found good training parameters.
FAQ
Comments (2)
While Civitai is in Low Resource Mode, you'll have to download the model and use off-site generation, apparently. It doesn't seem to be able to load this 64-dimension model right now.
If folks do use it, let me know what tags you've tried that don't work, what it does do well, etc. I've made some major improvements to my tagging software's export logic and right now I'm going through and retagging the training set to match what Mintygryph is trying that I didn't tag. More input gives me more to work with!
The Stable Diffusion version is the same size as the Illustrious version and it works fine, and the Illustrious version was working fine earlier. I've deleted and reuploaded it to see if that helps, but it looks like it didn't. It works fine locally. no idea what is going wrong. I've been working on major improvements to the data set and should have a V2 up in training soon anyway.