Type | |
Stats | 72 4,436 1.8k |
Reviews | (23) |
Published | Nov 9, 2024 |
Base Model | |
Training | Steps: 864 Epochs: 20 |
Usage Tips | Clip Skip: 2 Strength: 0.8 |
Hash | AutoV2 8CB2A2CDA8 |
(Note: This LoRA was not trained with photos. If you want a model for photo-realism, you may have to try elsewhere. Or you can experiment, it might work!)
I read somewhere that for making concept and style models like this, less is more, and to try to work with 20ish carefully curated and carefully tagged images. So I tried taking some of my big favorites, carefully cropping and cleaning them up, and hand-tagging each one.
I'll do separate sets of "looner" models based on specific poses or actions involving balloons pulled from my hoard. This is specifically "balloon blowing", not "pumping" or "inflating" or "filling" or any other variation. Don't let that stop you from experimenting, though!
I find best results with 20-30 steps, Euler Ancestral, CFG 6.5 (really, anything from 4 to 7 seems good, but I've had the most consistency with 6.5.).
Some "enhancement options" if you like experimenting or if you have ComfyUI and want to let it randomize a bit for you. (You'll see me using this trick a lot in my ComfyUI workflows. They should be embedded in the example images. Try it!):
{
giant balloon| big balloon | medium balloon| small balloon|
},
{shine|},
{necked balloon|},
{puffed cheeks|},
{star-print balloon|},
{doubled outline, white outline,|},
{from front|from side| from back|from behind| from below| from above|three-fourths view|dynamic angle|close-up|cropped, croch-pov|},