CivArchive
    ← All articles
    Published January 31, 2025by gl4mdiva

    Creating Ai FUR (vegan but realistic looking)

    169 views6 reactions1 comments on CivitAI3 collected
    fuzzyfur coatfurfuzzgeneration guide

    Those who have tried should have noticed that creating good-looking AI-generated fur is quite difficult.

    The main issue is that the word "fur" is rather abstract, unlike "stone" or "chair."

    Every text-to-image AI uses your input words via a prompt to generate a probability distribution for the image you want to create.

    The model you use is trained based on expectations of what each word should generate—determined by the model's trainers.

    Everything fluffy with hair is classified as "fur"—this includes fake fur, real fur, and fleece.

    Knowing this, the "fur" of a "fur coat" will never look great if you use "fur" as your only reference.

    You can try to work around this and increase alignment probability—the likelihood of achieving a more accurate result—by adjusting different model variables.

    A. Standard variables to consider:

    • Checkpoint – Some models perform better than others depending on their training data.

    • Prompt – Using terms like "angora rabbit fur" or "fox fur" is more effective because referencing the entire animal makes the description less abstract than just "fur."

    B. Alternative variables to consider:

    1. LoRAs – If a specific type of fur is trained into a LoRA, this is a game-changer.

      • If you want realistic results, achieving them without a LoRA is unlikely (as of 2025).

    2. IPAdapter models – Some IPAdapters act like a "one-image LoRA." You can input a picture of fur and get something similar as output.

    By adjusting these variables, you can significantly increase alignment probability, ensuring that the generated fur is closer to your expectations.

    The best way to fine-tune these settings is by using ComfyUI, as it allows for precise control over all parameters. Additionally, its flexible workflow system enables the creation of custom setups, allowing users to define new parameters that further increase alignment probability, improving the chances of generating realistic-looking vegan fur.

    (The cover image is sdxl base model, gl4mdiva fur coat lora, ipadapter, ComfyUI)