v2.0

Trained on 869 crops from the original movie with Danbooru style tags.

12 repeats, 10 epochs, batch size 12 - 869*12*10/12=8690 steps total.

I decided to use a unique jdhpps token as a trigger word to prevent the model from mixing it's knowledge of the character with my data. I tried using Judy Hopps first, but it seems to be hard to retrain the model from what it already knows.

LoRA supports all character's outfits from the movie:

blue bodysuit, zpd emblem (bare shoulders)

dress uniform, tie, shoulder cord, utility belt

casual clothes, grey t-shirt, bra strap, dark grey pants

police uniform, zpd badge, utility belt

meter maid uniform, orange vest, female police hat

farmer clothes, plaid shirt, rolled up sleeves, straw hat, jeans

Default negative prompt:

eye reflection, eyes reflection, blurry, worst quality, bad quality, lowres, malformed, deformed

LoRA also supports:

ears up, ears down, bunny teeth, glare in eyes, holding carrot pen, holding folder, holding notepad, holding photo

And all emotions from the movie in Danbooru style, including

:3, :<, :/, :d, :o

I recommend using dmpp_2m (DPM++ 2M) sampler with karras scheduler and 20-40 steps, without hi-res fix - looks much more detailed then euler a.

Hi-res fix with 0.4-0.5 denoise and dmpp_2m (DPM++ 2M) sampler, karras scheduler makes it even more detailed.

I personally recommend LoRA strength 0.7-0.8 (higher values can produce artifacts and similar backgrounds), but feel free to find better option for you!

Trained on https://civitai.com/models/3671?modelVersionId=187??92, and I recommend using this model.

Share your arts and don't forget to leave a review!

v1.0

Trained on 877 crops from the original movie with Danbooru style tags.

18 repeats, 2 epochs, batch size 10 - 877*18*2/10=3158 steps total.

I decided to use a unique jdhpps token as a trigger word to prevent the model from mixing it's knowledge of the character with my data. I tried using Judy Hopps first, but it seems to be hard to retrain the model from what it already knows.

LoRA supports all character's outfits from the movie:

blue bodysuit (bare shoulders, two tone bodysuit, dark pants)

dress uniform, tie, shoulder cord, utility belt

casual clothes, grey t-shirt, dark grey pants, bra strap, wristwatch (bracelet)

police uniform, zpd badge, utility belt (hand pads, knee pads)

meter maid uniform, orange vest, female police hat (ears down, hand pads, knee pads)

farmer clothes, plaid shirt, rolled up sleeves, straw hat, jeans

Add these tags for better character appearance and details:

detailed eyes, realistic eyes, purple eyes, large eyes, large ears, detailed fur, fluffy fur, two tone fur, sharp fur, detailed texture, female bunny, anthro, 4 fingers

Add these tags for realistic lighting and movie-accurate style:

new, 4k, best quality, high quality, masterpiece, detailed, realistic, photorealism, hdr, high contrast, vibrant colors, cinematic lighting, volumetric light

Default negative prompt:

eye reflection, eyes reflection, bad ???????, deformed iris, extra glint, low contrast, blurry, oversaturated, worst quality, bad quality, malformed, deformed, bad eyes, blurry eyes, cartoon, anime, 2d, flat, lack of details, simple, furry, human, text, artist name, tiny, short

LoRA also supports:

ears up, ears down, bunny teeth, eye reflection (eyes reflection), glare in eyes, holding carrot pen, holding folder, holding notepad, holding photo

And all emotions from the movie in Danbooru style, including

:3, :<, :/, :d, :o

Use DPM++ 2M Karras (sampler dmpp_2m scheduler karras) without hi-res fix

Use euler a Karras (sampler euler_ancestral scheduler karras) with hi-res (same seed!) using DPM++ 2M Karras and denoise from 0.3 to 0.6 - adds more sharpness and realism. euler a Karras is much better at eyes, teeth and clothes, but the fur will be soft and less detailed.

You CAN actually use DPM++ 2M Karras for both but it seems to mess up with glare in eyes. Hi-res might fix it or might not, keep it in mind. But the image definately will be more detailed.

I personally recommend LoRA strength 0.8-0.85, but feel free to find better option for you!

Trained on https://civitai.com/models/3671?modelVersionId=187??92, and I recommend using this model.

Share your arts and don't forget to leave a review!