[SD 1.5] Pokemon - Serena (Game): v3: character_pokemon_serena_v3.safetensors

SD 1.5 Standard Model
Download

There are detailed changes below the next paragraph, you might need to expand this version changes box!

I learned a few new things about creating LoRAs and wanted to see how far I can push my existing LoRAs without changes to their datasets (well... not quite for this one, see below). So there's still no support for alternate outfits, colors etc. for now, sorry! There are other LoRAs if you want those things. Not sure if I'll improve the LoRAs more than that in the future or make new ones but who knows.

LoRA Changes:

  • Now uses the standard "serena \(pokemon\)" instead of a custom trigger word to make things easier for everyone (please note though that the LoRA itself is still named "character_pokemon_cynthia_vX" as before!)

  • Switch to 32 dims (now trained directly instead of resizing from 128)

  • Switch to target weight 1 (increased from 0.7)

  • Changed a lot of training settings (too many to list, check LoRA metadata if you are interested) and updated to much newer scripts version (~3 months worth of changes)

  • I had to remove one images from the dataset as they confused the new training too much. As a result, some details for shots with looking behind might sadly become a bit worse in general and especially for close-ups. Hopefully not much else gets worse.

Example image changes:

  • Use CLIP skip 2 instead of 1 as I accidentally did before

  • Change settings for hir-res fix: Switch upscaler from None to Lanczos and increase denoise from 0.1 to 0.4 to get better results while still not changing TOO much

  • Use a negative embedding for the majority of negative prompting

Recommended weight: 1

Training model: Anything V3

Example image generation model: AbyssOrangeMix2 - Hardcore

Please read through the entire description (might need to be expanded) and the version change notes as they cover a lot of information about basic use cases and limitations. The description is always focused on the latest version. Thank you!

This is an SD 1.5 LoRA for the character Serena from Pokemon. It works for models based on SD 1.x only, it will not work for models based on SD 2.x or SDXL!

Example images were picked from pure 512x512 txt2img results and then re-created at 1024x1024 using txt2img with moderate Hi-res Fix settings (upscaler Lanczos, denoising strength 0.4). This improves faces and other details that are almost impossible to get correctly and consistently at 512x512 (limitation of the technology) while still giving a realistic impression of what it looks like. Pure 512x512 results will have more distorted faces and less detail.

Please see the version change notes for the training and example image generation models as well as the used weights as they might change between versions. Remember that you might need to adjust weights to best suit your use case!

Remember to add the tag serena \(pokemon\) (with backslashes intact) to your positive prompt.

The training set contained multiple variations of Serena's signature look (from the game, not the anime, sorry) so you might need to put combinations of the following tags in your positive or negative prompts to get the desired results:

  • hat

  • eyewear on headwear

  • sleeveless shirt

  • high-waist skirt

  • thighhighs

  • shoes

  • mega ring

  • shoulder bag

As the training set contained a few images with floating hair, you might need to add floating hair to the negative prompts if not desired.

Known Limitations / Problems:

  • This model is only trained on the game design for Serena, not the anime. I'm currently not interested to do the anime design, sorry.

  • The shoes and bag were not present in a lot of images and not a focus for the training so they sadly will not be consistent at all.

  • The face, eyes and mega ring will similarly be very inconsistent because of a lot of different interpretations in the training set, especially for wider shots. The eye color should be specified explicitly in the positive prompts, e.g. grey eyes.

  • Depending on the image composition (angle, distance etc.), it might be more or less difficult to remove the hat. Both hat and eyewear on headwear in the negative prompts and no headwear in the positive prompts should help but you still may need some luck.

  • Even with floating hair in the negative prompts, it tends to go a bit wild. To be fair, that's true for many images I've seen as well.

  • Shots from behind are very inconsistent due to extremely limited training data. Might for instance generate wrong hair, not apply poses correctly etc. Not much I can do without a bit more training data.