SD 1.5 vs. SDXL – Which Works Best for Japanese Spitz Fur Inpainting?

In this article, we’re exploring different Stable Diffusion checkpoints for inpainting, specifically SD 1.5 and SDXL. Our ultimate goal is to restore and enhance black-and-white Japanese Spitz fur. 

But before we go into specialized fine-tuning methods, we need to see what the “out-of-the-box” solutions can do.

Why compare SD 1.5 and SDXL?

SD 1.5 and SDXL are both versions of Stable Diffusion, a popular open-source AI model for generating images from text descriptions.

• SD 1.5 is the more established version for inpainting. Many people have created fine-tuned models on top of it, and it typically works well for a variety of tasks.

• SDXL is newer, offering higher resolution and sometimes more coherent details. But inpainting support and fine-tuned models for SDXL are still catching up.

We’re testing both to see how they handle Spitz’ fur restoration without specialized training, and if it’s even possible to adapt existing fine-tunes for this task.

Current state of inpainting models

You might wonder: “If we don’t like SD 1.5 or SDXL, why not just grab a different fine-tuned inpainting model?”

The unfortunate reality is that, for inpainting specifically, there aren’t many robust, well-documented fine-tuned models focused on detailed animal fur, particularly black-and-white fur.

Some specialized models exist on platforms like Civitai or Hugging Face, but they often target color portrait inpainting or other niche aesthetics.

Adapting them for black-and-white Japanese Spitz fur often requires re-training or complex prompt engineering that still yields mediocre results.

Bottom line: If you want top-notch results for a specific inpainting task (like restoring Spitz fur), you likely have to fine-tune or adapt the model yourself.

Technical observations: SD 1.5 vs. SDXL

Why are the results sometimes poor when adapting?

SDXL’s architecture differs significantly from SD 1.5, so a fine-tune or “adapter” designed for 1.5 won’t necessarily work well on SDXL.

Plus, most current fine-tunes are trained on color data with typical lighting conditions. Black-and-white Spitz fur is a narrow domain, so you need a specialized dataset to keep details sharp.

And, because SDXL is newer, the community has fewer inpainting success stories or step-by-step guides for problem-solving.

Our initial decision flow for now

Why SDXL’s infrastructure stands out

Despite the current lack of specialized inpainting fine-tunes, SDXL brings a unique two-container (or dual-UNet) architecture to the table. This approach separates certain stages of image generation – often handling coarse features in one phase and fine details in another. For large or detailed tasks like fur inpainting, this can be a game-changer:

Dual-stage denoising

SDXL’s design can process images at a higher resolution internally, maintaining more structural fidelity. When restoring fine textures like fur, more preserved structure means fewer artifacts.

Efficient memory usage

Although SDXL can be more VRAM-hungry overall, its modular approach allows advanced users to split or distribute the load across multiple GPUs or processes, scaling more smoothly.

In practical terms, if you have the hardware (e.g., a multi-GPU setup), SDXL can handle bigger images or more complex inpainting tasks without significant slowdowns.

Better text encoder & prompt handling

SDXL often ships with a more advanced text encoder stack. For tasks where you prompt “Restoring Japanese Spitz fur detail,” it can grasp context more intuitively if specialized fine-tuning is eventually done.

This advanced encoding helps when you combine textual instructions (like “enhance soft fur around the ears”) with an image mask. Though in practice, we still rely heavily on a good training/fine-tuning pipeline.

Conclusions & next steps

Both SD 1.5 and SDXL can inpaint black-and-white Spitz fur, but each has trade-offs. There’s no perfect “plug-and-play” solution without further fine-tuning.

We need a stable foundation, most likely SDXL, and a method to fine-tune or enhance fur details (hint: LoRA).

Our next article will show you why cropping and resizing can ruin fine details, and how our padding strategy fixes those issues to maintain perfect before/after image alignment.

Stay tuned so you can learn how to keep your inpainted image exactly the same size (and aspect ratio) as the original – without losing quality.