Tags of Note
Trigger: kpossible,
Style: flat color, narrow waist,
Clothing: black crop top, turtleneck, green cargo pants, black gloves, utility belt, belt pouch,
Mouth: closed mouth, open mouth, frown, smile,
Hands: clenched hand, clenched hands, arm support, hands on own hips, own hands together, hand up, hands up,
Eyes: half-closed eyes, closed eyes, looking to the side,
Legs: legs apart, legs together, kneeling, knees up, hugging own legs
Negative: eyelashes, shiny skin
If you enjoyed this LoRA, consider giving it aš
Description
Easier prompting, closer likeness compared to previous model.
FAQ
Comments (11)
How's the Jenny 2.0 version going? Will it also be released?
I fell into the trap of thinking more images = better. I did a lot of work and got worse results, lol. So, the dataset needs to be re-curated and re-balanced.
No plans on revisiting Jenny soon unless I can find a high quality source of the show (everything available on the internet is 480p), that would be a big help towards making a better model.
@saransaranĀ More images DOES equal better, what you don't need is excessive amount of steps. Too many steps like over 900 can make loras way too strong in the art style used in the reference images and thus more inflexible. And yes quality images are important but more important is the tagging, if you want your character and art style used to be separable you should tag the images with tags of what they are such as "anime screencap, anime coloring, toon \(style\)" etc. I can give you some tips if you're interested.
@BecauseNoOneElseWillĀ More high quality images = better, yes. But quality matters much more than quantity. A LoRA trained with 20 high quality images will outperform a LoRA with 20 high quality images + 20 low quality images. At least, that's what I've found with testing. Lower quality images have a really negative impact. For every low quality image, I need like 9 high quality images to undo the damage it causes when training.
You're certainly right about over training! I save every 50 steps and test each one to see where the training is over fitting vs under fitting.
If you have any tips or feedback, feel free to DM me.
@saransaranĀ Couldn't you just use AI upscaling software and additional cleanup in Photoshop to get the high-def images you need?
@Da_Cartoon_HootĀ I haven't had any success with upscalers. They always introduce errors into the image, like smearing small features, creating color bleed, or inventing undesired details. So they seem to create more problems than they solve.
I did get some success with doing a very gentle img2img pass, then importing into Photoshop to apply a highpass filter to eliminate some of the blurriness, and correct minor mistakes created by the img2img pass. But even that has its limits. It's a very time consuming and effortful process for one image, and doing that for the hundreds of images in Jenny's dataset just isn't feasible. The best option is to start over and collect data from a high quality recording of the show (which there doesn't seem to be any).
@Da_Cartoon_HootĀ From my personal experience, a well trained lora with quality images, good prompting and a decent amount of steps you won't have any need for extra work on upscalers and such.
Any Compliance Chips in the training data?
No, the data was collected in a broad sweep, so no episode specific props are tagged or included. That' a good idea for a future version, though!
Nicely Done!
Thank you! š„°
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.