README.md

Styles Used:

1. Oil style
2. Xyz
3. Allante
4. Moebius
5. Polygons

Result of Experiments with different styles:

Prompt: "a cat and dog in the style of cs"
"cs" in the prompt refers to "custom style" whose embedding is replaced by each of the concept embeddings shown below

---

Prompt: "dolphin swimming on Mars in the style of cs"

Result of Experiments with Guidance loss functions:

Prompt: "a mouse in the style of cs" Loss Function: python def loss_fn(images): return images.mean()

def loss_fn(images):
    return -images.median()/3

def loss_fn(images):
    error = (images - images.min()) / 255*(images.max() - images.min())
    return error.mean()

Prompt: "angry german shephard in the style of cs"

def loss_fn(images):
    error1 = torch.abs(images[:, 0] - 0.9)
    error2 = torch.abs(images[:, 1] - 0.9)
    error3 = torch.abs(images[:, 2] - 0.9)
    return (
        torch.sin(error1.mean()) + torch.sin(error2.mean()) + torch.sin(error3.mean())
    ) / 3

---

Prompt: "A campfire (oil on canvas)"

def loss_fn(images):
    error1 = torch.abs(images[:, 0] - 0.9)
    error2 = torch.abs(images[:, 1] - 0.9)
    error3 = torch.abs(images[:, 2] - 0.9)
    return (
        torch.sin((error1 * error2 * error3)).mean()
        + torch.cos((error1 * error2 * error3)).mean()
    )

---

def loss_fn(images):
    error1 = torch.abs(images[:, 0] - 0.9)
    error2 = torch.abs(images[:, 1] - 0.9)
    error3 = torch.abs(images[:, 2] - 0.9)
    return (
        torch.sin(error1.mean()) + torch.sin(error2.mean()) + torch.sin(error3.mean())
    ) / 3