Aesthetic Quality Modifiers - Masterpiece
Training data is a subset of all my manually rated datasets with the quality/aesthetic modifiers, including only the masterpiece tagged images.
ℹ️ LoRA work best when applied to the base models on which they are trained. Please read the About This Version on the appropriate base models, trigger usage, and workflow/training information.
Recommended prompt structure:
Positive prompt (quality tags at the start of prompt):
masterpiece, best quality, very aesthetic, {{tags}}, {{natural language}}Description
Trained on Krea 2 Raw base with diffusion-pipe.
Strength of 1.5 is recommended when applied to the Turbo model.
Previews generated with a slightly adjusted int8 quant workflow by Winnougan: https://huggingface.co/Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8
Workflow is in included in the preview images, you can drag and drop into ComfyUI.
Same dataset as v5.1, adjusted captions to simply use "masterpiece, very aesthetic" trigger tags + NL captions for Krea 2 - which is assumed to have no booru tags knowledge
Can benefit from being used together with the krea2filterbypass LoRA:
https://civitai.red/models/2728234/krea2filterbypass?modelVersionId=3067151
See: https://civitai.red/images/135097274
Training configs:
https://github.com/motimalu/diffusion-training-configs/tree/main/diffusion-pipe/krea/configs
FAQ
Comments (12)
All anime dataset?
For the most part yes, it might depend on how broad your definition of anime is.
Does this lora replace the score things? im confused :D
Hello,
If a base model uses some form of score tags, you can still use them.
These are just LoRA that used "masterpiece" and "very aesthetic" tagged images, you could consider them like any other style LoRA with trigger tags.
Not really aiming to do anything to other tags the models are trained with, for example no distillation techniques that force a model to always express some tags effects unprompted are used.
Score tags can interfere with the expression of many styles, so omitting them may allow this LoRA to be expressed a little better.
They can also help stabilize generations if the base models used them extensively though, so it is up to your preference whether or not to use them.
@motimalu Thanks a lot for ur great explanation, really appreciate it :)
pls make a version for chroma
Are there any recommended intensity parameters?
For models with a Raw/Turbo split where the LoRA is trained on the Raw base and applied to the Turbo distillation (Krea 2 Turbo, Z Image Turbo etc), a higher strength of ~1.5 helps when applied to the Turbo model.
Applied to a base other than what this was fine-tuned against, like a merge, also benefits from increasing the strength slightly too.
Basically it probably wont work as well when applied to models that diverge significantly from the base that it was trained on - but you can try increasing the LoRA strength in those cases.
Raw/Turbo models don't have a good way around this AFAIK, maybe "Quantization-Aware Training" but ┐(´ー｀)┌
@motimalu ty
Do the CLIP strength and LoRA strength need to be kept the same when using a LoRA? Thanks.
@zjc1772665101270 No need to stay consistent.
amazing！