Sign In

Experimental ☀️ from Huggy

183
3.2k
22k
101
Verified:
SafeTensor
Type
LoRA
Stats
40
292
Reviews
Published
Sep 29, 2024
Base Model
SDXL 1.0
Training
Steps: 1,328
Epochs: 9
Usage Tips
Clip Skip: 1
Strength: 1
Trigger Words
aaAR7ip
oil painting
(wind blown:2)
Hash
AutoV2
53E0C350C3
default creator card background decoration
Second Birthday Badge
huggy's Avatar
huggy

You guys, We’ve all seen how some users, like DaOneWhoDontSnooze, are taking advantage of the reporting and rewards system over at Civitai. If you check out their profile, you’ll notice they’ve got a “Guardians” badge—supposedly there to protect the community. On average he reports 1,000 artworks every month. But wait, take a closer look at their uploads. Some of the content is seriously pedophilia. Isn’t it wild that someone like this is being allowed to ‘keep us safe’ when they’re clearly abusing the platform? Something has got to change.

📒 This is a big lora complex, read following table!

📒 Everything is an work in progress:

As you see, I uploaded lots of similar lora made with different epochs, LR and resource. Every lora is subtly unique as I will systematically compare one another by its quality (over sharpen / glitch / half baked), resemblance (to training resources) and character features (torso deformation, poor eyes, unnatural looks) in the future.

📒 Useful Tips:

  • Use the Lora with epoch below 7 for a stable result, like V03E06, though lora with higher epoch like V03E10 has better resemblances to the training source;

  • some loras include multiple trigger codes (aaXXXxx, aaXXXxx, etc), representing different styles / characters.

  • USE fewer triggering words each time for best result;

  • Use CFG between 4.5 (complex scene) to 7.5 (simple);

  • For anime start training steps from a lower value like 19 with Euler a sampler, if you are satisfied with the result, use hi-res fix / face fix (and set Euler to 29) to increase over all quality;

  • For realistic photos, use DPM++2M Karra, 30+ steps for close up portrait, 35+ steps for full body shot.

  • There's no universal prompt / lora weight for every generation. If you're using anime style / character lora with realistic checkpoint, starting with a lower lora weight like 0.7. Ignore the lora weight in sample images, due to a recent bug, image generated online will display lora weight of 1 regardless the actual value.

  • step number setting (which is the weight for all tags) is closely associated with the complexity of the prompt, if your outcome has extra objects like fingers or limbs, try lower the step value (and the lora weight too); if your outcome has objects disappeared, lack of quality, texture and reality, try raising the step value.

  • Compose a prompt and wait for the generation seems like flipping through a mystery novel, you never know what twist is coming next. It's a way to explore the boundless of AI model. To get the exact outcome of what you planned is difficult as you have to find the proper glossary for each tag, and each word itself is a vague variant with various synonyms.

  • The training files are mostly focus on solo character, which means if you intent to generate a image with couple, you might get unexpected result. To fix that you need add additional lora which is trained with multiple people (like ExpressiveH) or a pose lora. A good multiple people lora should contain as many pose as it can with couple, threesome or a group of people and it is definitely a tough job.

📒 Issues to be fixed in the next version:

1) training base on Pony get versatile results than training base on SDXL, but when prompts contain 1 human and 1 animal, Pony has a great tendency of blending two species, put source_pony, source_furry in the negative prompts won't help. Does anyone knows how to solve this?

2) reduce training source quantity down to ??? or less, as this is a common problem across most of my loras, too many training images reduce the outcome quality, quality of the source images are more important.

3) Increase training source resolution.

4) Reduce training steps down to 1000, as I found lower training epochs lora got stable results

📒 Checkpoint Rating for Illustra V04E03

This is a lora complex which contains 11 (100 is my goal) unique styles and poses. Currently I have a list of ideas to add into this lora complex. It takes enormous time for me to prepare source images. Each Lora will need 2-15 iterations to fix glitches. I'd like to test my lora with all available checkpoints online before publishing, but I cannot do any of these since my buzz is running oooout, donate buzz if you like what you see.