Read the "About this version" section to know what changes I made to the model.
TL;DR: This model is a direct fine-tuning from AnyLoRA, unlike AingDiffusion, which merges multiple models and some personal datasets.
The dataset used to fine-tune the model is still small, but already significantly affects the model.
The model is still in development, and I will significantly update the dataset.
===
ElixirProject is a direct fine-tuning from AnyLoRa. This model is capable of generating high-quality anime images.
Guide to generating good images with this model
Use any negative embeddings (I recommend my stuff, AissistV2, which I develop for my models.)
The recommended sampler are "DPM++ 2M Karras" and "DPM++ 2M SDE Karras"
Hi-res fix is recommended. The recommended upscaler is "4x-AnimeSharp"
Set clip skip to 2 (recommended, fine-tune dataset is trained on clip skip 2), ENSD (eta noise seed delta) to 31337 [to replicate image], and eta (noise multiplier) for ancestral samplers to 0.667.
Description
I've retrained the dataset using the Prodigy scheduler, and the result seems better than v1.3. I'm also adding a new dataset to the model. So there are 2 datasets right now.
FAQ
Comments (2)
If you were training a character LoRA would you recommend using this model over AnyLoraCleanLinearMix?
And additionally are there any VAE requirements for this?
I haven't tried to train a LoRA with ElixirProject lol. But I think it should be no problem to train a LoRA using the model.
The VAE is baked into the model, you can override it, tho. The baked VAE is vae-ft-mse-840000-ema-pruned.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.