This is an Anima Edit LoRA designed for manga colorization. For inference, it must be used with the ComfyUI-Cosmos-Reference node.

The LoRA was originally intended to support reference-guided colorization. However, testing showed that its ability to follow the reference image is limited. The reference image mainly influences the overall color palette and color distribution, rather than reliably transferring specific colors to corresponding characters or regions.

The training dataset consisted of 2,566 paired images at 1024 resolution. The input images were augmented with various degradations, including blur, downscaling, and quality reduction. As a result, the LoRA can improve the visual quality of lower-quality input images to some extent while colorizing them.

For best results, the input image should not exceed approximately 2 megapixels. Colorization quality may degrade significantly at higher resolutions.

The dataset also included a small number of synthetically generated photorealistic images, so the LoRA can colorize real-world images to a limited extent as well.

The activation keyword is:

colorized

However, the LoRA can also be used without entering any prompt.

Colorization has currently only been tested with a Turbo LoRA setup, specifically Turbo-ANIMA-v4.