Colorfix lora for Anima Preview.
Adds color variety, removes blur from backgrounds, boosts prompt adhesion (mostly, not always), adds details uniformly.
Intended to decrease amount of same colors in image. As a side effect adds details a bit.
It was not trained on any common imagery, see my noob v-pred colorfix for explanation.
But it is a cancellation lora, so it brings up certain biases in model while cancelling others. Think of it as deletion of concept of one image having a uniform color. As a result it boosts shading.
I noticed general shift towards cartoonish style and painterly style. If you see that and don't want it - add toon \(style\), painterly, to the negative.
Designed to work around weight of 1. With additional loras or finetunes may need more. Use with negative weight to enforce less details and more uniform color.
Details are a side effect, this is not a proper detailer model. At least I think so now.
Description
Trained on Anima Base. The model rarely needs it, mostly on overtrained artists. Sometimes breaks text and hands, sometimes fixes them. Really follows unstable nature of the model itself.
FAQ
Comments (7)
I don't know how it does its magic, but it does magic.
I don't like having "default go-to LORA's" but here we are. It's perfect.
It adds jank at times, let's be honest. Hope I will find all the tooling to make proper detailer
How did you train this Lora? Can you explain in detail
Dataset and comments here: https://civitai.red/models/1555532/noob-ai-xl-v-pred-stoopid-colorfix-toolset?modelVersionId=1760239
Basically I am just leveraging the fact that ADAMW trains big features first, details later, so I trained "general concept" of undertrained flat color then cancelled it. This is a cancellation lora in essence, and pretty basic.
@lse14 btw where did you train sliders for anima? I urgently need those for proper detailer. Implementation in ostris is ready?
@Volnovik I obtained these detail sliders using a custom script with overfitted delta training on the base model.The slider LoRA was trained on a base model that had been merged with an overfitted base LoRA.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.