In our previous article, we established that combining LoRA and Textual Inversion provides the optimal foundation for inpainting black-and-white Spitz fur. Now, we get to prompt engineering to craft the perfect prompt and parameters for generating highly detailed, realistic fur.
Below, we present a series of experiments, ordered from least to most effective, based on the results we got. Let’s explore what worked, what didn’t, and how we arrived at the ultimate setup.
Experiment 1: The simple prompt
We started with a basic prompt to establish a baseline.
- Prompt: “detailed fur of a Japanese Spitz”
- Negative Prompt: “”
- Strength: 0.2
- Guidance Scale: 7.5
👉 What we got: The model generated detailed fur, using LoRA and Textual Inversion. However, the result was basic and generic, lacking the “super detailed” quality we aimed for. The fur missed the fine textures and unique traits of a Japanese Spitz’s coat, making it a functional but underwhelming starting point.
Experiment 2: Adding descriptive flair
Next, we enriched the prompt with vivid adjectives to boost detail.
- Prompt: “super detailed, high-quality fur of a Japanese Spitz, with fine textures and realistic shading”
- Negative Prompt: “”
- Strength: 0.2
- Guidance Scale: 7.5
👉 What we got: This method produced slightly better results than the previous experiment – but the pictures looked almost the same. The fur gained depth and refinement, with subtle shading adding realism. While it was an improvement, it still fell short of the precision and detail we wanted.
Experiment 3: Leveraging the Textual Inversion token
We then incorporated our custom Textual Inversion token, <dog_fur_style>, to use the specific fur style we trained.
- Prompt: “A fluffy dog’s <dog_fur_style> fur. Keep original shape.”
- Negative Prompt: “”
- Strength: 0.2
- Guidance Scale: 7.5
👉 What we got: The result was noticeably better than the previous attempt. The <dog_fur_style> token aligned the fur with our trained high-quality style, enhancing consistency and texture. This tailored approach brought us closer to our goal.
Experiment 4: Introducing a negative prompt
To refine the output further, we added a negative prompt to avoid common pitfalls.
- Prompt: “<dog_fur_style> super detailed fur of a Japanese Spitz”
- Negative Prompt: “blurry, low quality, artifacts”
- Strength: 0.2
- Guidance Scale: 7.5
👉 What we got: This method gave us a much better outcome than before. The negative prompt reduced blurriness and artifacts, resulting in cleaner, crisper fur details. The quality improved slightly, but without altering the core prompt, proving the value of this addition.
Experiment 5: Boosting strength
We increased the strength parameter to see if it would enhance the details further.
- Prompt: “A Japanese Spitz <dog_fur_style> with super detailed fur”
- Negative Prompt: “blurry, low quality, artifacts”
- Strength: 0.5
- Guidance Scale: 7.5
👉 What we got: The result was much better than the previous experiment. The higher strength (0.5) amplified the fur details, making them richer and more pronounced. However, slight over-modification appeared in some areas, hinting that we needed to balance strength carefully.
💡 Highlighted insight: Higher strength improves fur quality but also increases processing time. While it enhances details, it can introduce minor artifacts if not managed properly.
Experiment 6: Even more strength
Curious about higher strength, we tested values of 0.7 and 1.0.
- Prompt: “A Japanese Spitz <dog_fur_style> with super detailed fur”
- Negative Prompt: “blurry, low quality, artifacts”
- Strength: 0.7 and 1.0
- Guidance Scale: 7.5
👉 What we got: At 0.7, the fur became even more detailed and slightly better than Experiment 5, though minor artifacts emerged around the edges.
At 1.0, the result took a downturn – while details peaked, prominent anomalies and artifacts appeared, making the fur look unnatural. This showed that strength at 1.0 was too extreme for our goal.
💡 Highlighted insight: High strength (e.g., 1.0) significantly increases processing time and boosts fur detail, but it also generates artifacts. Zoning fur areas (excluding legs, etc.) helps reduce these issues, keeping the focus on the fur itself.
Experiment 7: Amplifying guidance scale
We adjusted the guidance scale to test its impact on prompt adherence.
- Prompt: “A Japanese Spitz <dog_fur_style> with super detailed fur”
- Negative Prompt: “blurry, low quality, artifacts”
- Strength: 0.2
- Guidance Scale: 10.0
👉 What we got: This method was slightly better than Experiment 4 (same strength, lower guidance). The higher guidance scale (10.0) sharpened the fur details by making the model focus more on the prompt. However, a subtle stylization crept in, slightly reducing the natural look we wanted.
💡 Highlighted insight: A guidance scale of 10.0 enhances detail, but needs careful tuning. Pairing it with higher strength can balance artifacts, especially when zoning fur parts for excellent results.
Experiment 8: Pushing the limits with a high guidance scale
For a bold twist, we tested an unusually high guidance scale – well beyond the typical range – to see how far we could push the model.
- Prompt: “A Japanese Spitz <dog_fur_style> with super detailed fur”
- Negative Prompt: “blurry, low quality, artifacts”
- Strength: 0.2
- Guidance Scale: 15.0
👉 What we got: This method produced the worst result of all our experiments. The guidance scale of 15.0, an anomaly parameter compared to the standard 2 or our base 7.5, hyper-focused the model on the prompt to a fault. While it generated ultra-detailed fur, the texture became excessively stylized, with over-the-top contrasts and unrealistic patterns that deviated far from natural Spitz fur. This wild card proved too extreme, sacrificing quality for exaggerated effects.
Experiment 8: Combining the best
Finally, we combined the most effective elements into one setup.
- Prompt: “A Japanese Spitz <dog_fur_style> with super detailed fur”
- Negative Prompt: “blurry, low quality, artifacts, unnatural colors”
- Strength: 0.7
- Guidance Scale: 10.0
👉 What we got: This method delivered the best result of all. The fur was highly detailed, natural-looking, and nearly artifact-free – far superior to all previous experiments. The strength of 0.7 and guidance scale of 10.0 struck a perfect balance, especially when we zoned in on fur areas (body only, excluding legs), minimizing anomalies and ensuring seamless inpainting.
💡 Highlighted insight: A strength of 0.7 with a guidance scale of 10.0 maximizes fur quality while keeping artifacts in check. Processing time increases, but zoning fur parts (avoiding legs, etc.) eliminates most artifacts, making this setup excellent for Spitz fur.
The perfect prompt for Spitz fur inpainting
After extensive testing, we’ve settled on the ultimate prompt setup for inpainting black-and-white Spitz fur:
- Prompt: “A Japanese Spitz <dog_fur_style> with super detailed fur”
- Negative Prompt: “blurry, low quality, artifacts, unnatural colors”
- Strength: 0.7
- Guidance Scale: 10.0
This configuration, paired with our LoRA and Textual Inversion stack, produces the most detailed, realistic, and clean fur inpainting results. By balancing strength and guidance scale – and zoning fur areas – we’ve achieved exceptional quality.
Leave a Reply