A LORA for Honey Lemon from Big Hero 6. This LORA was trained to work with PonyXL based models, so use the standard prompts: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up
I'm still not completely happy with this, so you'll probably see a new version soon.....but it can produce some decent results. I'd recommend .7 strength. ⚠️ NOTE: this was not trained on Honey Lemon's "superhero" outfit. That will be a separate version.
Description
⚠️ NOTE: this was not trained on Honey Lemon's "superhero" outfit.
FAQ
Comments (9)
I hope you do a new version, it needs a lot of work. Most of my results trying to replicate the D*sney style are really failing.
I completely understand this. For some reason that I still don't completely understand, I've struggled more with training Honey-Lemon than almost any other character. I've actually deleted a couple of previous versions because I just wasn't happy with them. Maybe it's time i give it another go?
@tomorrowlandbailey937 Some characters are finnicky like that for sure. Tried training an IRL gravure model and SD HATED trying to train her lol.
I discovered that because I am using a Pony derivative (Godiva) the normal 0.85 LoRA strength was nowhere near enough. The new technique uses strength 1.3 for much better results.
Any updates on the way?
Possibly. For some reason that I still don't completely understand, I've struggled more with training Honey-Lemon than almost any other character. I've actually deleted a couple of previous versions because I just wasn't happy with them.
@tomorrowlandbailey937 Ah well. Appreciate the effort!
I used this in my comparative study article about lora strengths. If you'd like to see some examples of this lora being used with AutismMix and no other lora's, check it out.
Great article! You're correct, training this one was incredibly difficult. I generally use at least 150+ images to train my LoRAs......and as a side character from a single movie, there just aren't that many places to get images of her. The bulk of the data-set was cropped, frame by frame from scree captures. Most of these feature Honey in the background, meaning they where not the greatest quality even before I had to crop them. It took three different runs of image gathering, pruning and training before I got something I was finally (somewhat) happy with. I pride myself on training LoRAs that don't take a lot of work to generate decent images. For many of them you can get a reasonable image with a default "1" strength and basic prompt.......that's definitely not the case with this one. You really have to work to get images of Honey right.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.