Artificial Intelligence in Interior Design

Core AI Architecture

The core of our DEC-AI tool is built upon a robust integration of advanced diffusion models:

• Stable Diffusion v1.5 [3]: Serves as the base latent text-to-image diffusion model, capable of generating photorealistic images from textual descriptions. It provides the fundamental image generation capability.
• FreeU [6]: Enhances the U-Net architecture of the diffusion model by strategically re-weighting skip connections and backbone features. This "free lunch" mechanism improves generation quality and computational efficiency without compromising performance, particularly by optimizing feature flow with a Fourier filter.
• DiffEditor [2]: A novel model for fine-grained image editing, addressing issues of accuracy and flexibility. It introduces image prompts alongside text prompts, combines stochastic and ordinary differential equations for consistency, and incorporates regional score-based gradient guidance for precise modifications like object moving, resizing, and pasting.

This synergistic combination allows our tool to not only generate high-quality initial designs but also to perform precise, flexible, and context-aware edits.

Practical Design Capabilities

Our AI tool offers intuitive functionalities for interior design modification:

• Appearance Modulation: Seamlessly replaces the appearance of objects within the same category across different images. This involves sophisticated algorithms like color correction and style/shape adjustments to ensure the object blends naturally into the background image.
• Object Moving and Resizing: Allows users to easily reposition and adjust the size of objects. A scaling coefficient is applied to modify object dimensions, with zero-padding used to fill any newly created blank spaces, ensuring visual coherence.
• Object Pasting: Enables the transfer of objects from one image to another, ensuring the pasted object appears smooth, natural, and perfectly integrated into the new environment, adapting to the target image's aesthetics.

These features, powered by diffusion models, enable rapid prototyping and personalized design adjustments that are visually stunning and contextually appropriate.

Empirical Performance & Visual Quality

Our model was implemented using Stable Diffusion v1.5 with 1x NVIDIA RTX 4090 (24GB VRAM) and DDIM sampling (50 steps). We compared its performance against DragonDiffusion [1] across three key metrics: Cosine Similarity, SSIM (Structural Similarity Index), and PSNR (Peak Signal-to-Noise Ratio).

The results consistently show our model outperforming DragonDiffusion in "Moving and Resizing" and "Object Pasting" for most metrics, indicating superior visual quality and seamless integration. For "Appearance Modulation," the performance was comparable or slightly lower in SSIM/PSNR but similar in Cosine Similarity.

Visually, our model produced more realistic and cohesive designs, with smoother surfaces, better shading, and accurate object recreation, as evidenced in comparisons for object pasting (Figure 8), bedroom designs (Figure 9), and object movement (Figure 10).