Aesthetic Quality Modifiers - Complete
Training data is a subset of all my manally rated datasets with the quality/aesthetic modifiers, including only the masterpiece, best quality , or very aesthetic tagged images (7273 total).
Superset and final of models in the Aesthetic Quality Modifiers Collection.
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 and workflow/training information.
Recommended prompt structure:
Positive prompt (quality tags at the end of prompt):
{{tags}}
masterpiece, best quality, very aestheticGeneration Settings:
Previews are generated in Forge with upscaling and adetailer.
For Noobai V-Pred, a ComfyUI workflow reference with DynamicThresholding, Upscaling, and FaceDetailer can be found here: https://civarchive.com/posts/11457095
Description
Trained on NoobAI-XL (NAI-XL) V-Pred 1.0-Version
With the kohya_ss dev branch, v_paratemization and zero_terminal_ssr enabled.
Updated the training data to include some recently rated datasets
Total 6,661 images
Updated some tags to match noobai tagging conventions
very aesthetictags tovery awaby {artist}tags toartist:{artist}
Previews use CFG of 10 with DynamicThresholding (CFG-Fix) Integrated in Forge
A CFG of 5.5 is recommended if not using this, references on the effect of CFG on image saturation: https://civitai.com/posts/11143736
ComfyUI reference example with DynamicThresholding: https://civitai.com/images/49703093
FAQ
Comments (6)
What should the cfg for hires fix be for v-pred? I followed the dynamic settings on the green gem lora. Also what schedule type? I've been using ays
>What should the cfg for hires fix be for v-pred
I've been using the default of 5, but should match the cfg used for t2i if DynamicThresholding is enabled, since it will be applied to hires fix also
>Also what schedule type? I've been using ays
AYS gives me more saturated results, maybe playing with CFG it could be nice
NOOBAI/L_A_X have recommended to use DDIM on their model card:
2. Choice of sampling methods: Although NOOBAI XL-VPred 1.0 supports most sampling methods, V prediction does not support the Karras series of sampling. Therefore, we suggest using Euler and DDIM sampling methods for more stable outcomes.
awa is the name of an artist, with very awa you're slightly moving toward their style every time.
Thanks for pointing that out. That's true - though the decision to use very awa for the top 5% of images in terms of aesthetic score is made by the Noobai models some of these LoRA are trained on.
I thought the v2 v-pred version which used very awa instead of very aesthetic had some issues, so I might return to very aesthetic in the future.
Does the illustrious model only works with illustrious models? or can it be used with other models?
Hello, LoRA work best on the models they are trained on - though you could technically use these on any SDXL based models.
I'd recommend using the illustrious version with a model that has illustrious in its lineage at least though - either through further finetuning (noobai) or checkpoint merging (wai-nsfw) for example.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.